NASA Contractor Report 180852

Improved Numerical Methods

for Turbulent Viscous
Recirculating Flows
rSASA-CE-180852) IHPECVED KUBBBICfii METHODS
POB 10EBD1IH1 VISCOUS BECIBCOIM JNG FiOSS
Final Bepcrt
(Ivco-Evecett J&csearci Lab.)
2^ p
CSCL 01A

N88-25445

63/02

A. Turan and J.P. VanDoormaal

Avco Research Laboratory, Inc.
Everett, Massachusetts

July 1988

Prepared for
Lewis Research Center
Under Contract NAS3-24351

NASA
National Aeronautics and
Space Administration

Onclas
0154053

TABLE OF CONTENTS

Section
'\

Page
Foreword

1.0

SUMMARY

2.0

INTRODUCTION

2.1
2.2

Background
Objectives and Approach

2.3

Organization

3.0

3.1

Introduction

10

3.2

The Mathematical Problem

3.2.1 General Conservation Equations
3.2.2 Modelled Form of the Conservation Equations
Algebraic Representation of Conservation Equations

11
12
13
15

3.3.1

16

3.4

4.0

OVERVIEW OF THE COMPUTATIONAL DETAILS OF THE SOLUTION

ALGORITHMS FOR FLOW PROBLEMS

3.3

>,
*-*.,

:v

Domain Discretization

3.3.2 Discretization of Differential Equations

3.3.3 Discussion on the Discretization and Solution
Schemes
General Solution Methodologies for Incompressible
Flow Problems
3.4.1 Linearization of Mass and Momentum Conservation
3.4.2 Solution of the Unreduced System of Equations
3.4.3 Solution via a Generalized Segregated Approach

ASSESSMENT OF IMPROVED DISCRETIZATION SCHEMES AND SOLUTION

ALGORITHMS FOR GENERAL (INCOMPRESSIBLE) FLOW PROBLEMS

4.1
4.2
4.3

18
20
21
22
24
25
29

4.4

Introduction
Characteristics of Accurate Discretization Schemes
Computational Details of the Improved Solution Schemes
for Incompressible Flows
Selection of the Assessment Criteria and Methodology

29
30
33
37

4.5

Closure

40

Section
5.0

.,

Page

DESCRIPTION AND PRELIMINARY ASSESSMENT OF CANDIDATE SCHEMES . 41

5.1

Improved Discretization and Related Techniques

5.1.1 Second Order Upwind Differencing Scheme
5.1.2 Advanced Skewed Upstream Differencing Schemes
5.1.3 Higher Order Schemes Including Various
Compact Implicit Differencing Schemes.
5.1.4 Explicit Dissipation Schemes
5.1.5 Improved Flux Blending Schemes
5.1.6 Varlational Discretization Schemes Including
Finite Elements
.
.,
5.1.7 Spectral Methods
5.1.8 Lagrangian Methods

41
41
42
;,
43
45
48
49
51
53

5.1.9 Adaptive Grldding and Modified Equation

Analysis

5.2

5.3
5.4
6.0

Improved Solution Algorithms for Enhancing Convergence

and Efficiency of Methods for Incompressible Flows

56

5.2.1 Approximate Factorization Techniques

5.2.2 Conjugate Gradient Acceleration
5.2.3 Block Correction Acceleration

57
57
58

5.2.4 Multilevel Multigrid Acceleration

5.2.5 Coupled Numerical Schemes
Preliminary Evaluation of Candidate Techniques
Selection of Four Schemes for Further Quantitative
Evaluation

58
60
62

DERIVATION OF THE SELECTED SCHEMES FOR TWO-DIMENSIONAL

EVALUATION

6.1

6.2

6.3

54

Finite Volume Discretization and Solution of the

Conservation Equations for a Scalar, Momentum and Mass
6.1.1 Calculation and Discussion of Interface Fluxes
Second Order Upwind Differencing Schemes
6.2.1 Derivation and Characteristics of Finite
Difference Equations
6.2.2 Solution Details
Advanced Skewed Upstream Differencing Schemes
6.3.1 Influence Point Equations

65
66

66
70
75
75
80
80
80

6.3.2 Assembly of Influences and Discussion of

the Finite Difference Equations

ii

87

6.3.3

*
6.4
*

6.5

6.6
6.7
7.0

7.1

7.2

^
*
'

7.3

Mass Weighted Skewed Upstream Differencing

Scheme for Positive Definite Variables
89
6.3.4 Further Considerations in the Implementation of
Advanced Skewed Upstream Differencing Schemes
91
6.3.5 Boundary Conditions
.93
Compact Implicit Discretization Schemes
95
6.4.1 Derivative Compact Implicit" Scheme
95
6.4.2 Generalized Operator Compact Implicit Schemes
97
6.4.3 Control Volume Based Operator Compact
Implicit Method of Exponential Type
104
6.4.5 Multi-Dimensional Extensions of Operator
Compact Implicit Schemes
109
Improved Solvers for a General Segregated Solution
Algorithm
117
6.5.1 Implicit Base Solvers for Pressure/Correction
Equation of SIMPLE Derivative Algorithms
117
6.5.2 Conjugate Gradient Acceleration
119
6.5.3 Block (Additive) Correction Acceleration
122
6.5.4 Additive Correction Multigrid Acceleration
126
Assessment of Improved Solver Characteristics
130
Closure
, 131
DISCUSSION OF ONE AND TWO-DIMENSIONAL TEST CASES

132

Evaluation of improved Solver Performance for a

General Scalar (Pressure) Equation

132

7.1.1

132

Test Problems and Details of Implementation

7.1.2 Results of Numerical Experiments

134

7.1.3 Discussion of Results

Evaluation of Improved Pressure ~ Velocity Coupling
Algorithm and Solver Performance in the TEACH Code
7.2.1 Details of Implementation
7.2.2 Test Problems and Procedure
7.2.3 Results of Numerical Experiments
7.2.4 Summary and Conclusions
Evaluation of Improved Discretization Schemes
7.3.1 Second Order Upwind Differencing Scheme
7.3.2 Variants of Skewed Upstream Differencing
Schemes
7.3.3 Compact Implicit Discretization Schemes

137

in

139
139
141
143
146
147
148
156
171

Section

....
7.;4

8.0

Closure and Selection of Schemes for Three-Dimensional

Evaluation
.

8.3

Evaluation of Techniques for Improving Computational

Efficiency
8.1.1 Strongly Implicit Procedure vs Incomplete
Choleski
8.1.2 Additive Correction Multigrid (2) vs Additive
Correction Multigrid (3)
.
8.1.3 Block Correction
'
8.1.4 Conclusions
Considerations in Extending MW-SUDS And LP-Suds
Formulation In Three-Dimensions
8.2.1 Explicit LP-SUDS Integration Point Equation
8.2.2 Explicit MW-SUDS Integration Point Equations
8.2.3 Flux Element Assembly
Three-Dimensional Test Case

8.4

Closure

8.2

10.0

DISCUSSION OF THREE-DIMENSIONAL TEST CASE

8.1

9.0

Page
182
186

186
187
189
189
190
190
192
194
196
197
203

CLOSURE

204

9.1
9.2

204
208

Summary and Discussion

Recommendations for Future Research

REFERENCES

211

Appendices

A
B
C

Implementation of Modifications in the Pratt and Whitney

3-D TEACH Code to Improve Accuracy and Efficiency

218

Derivation of Coefficients.for R and Q Matrices 1n the

Classical Operator Compact Implicit Scheme

234

Details of Truncation Error Series for Classical

Operator Compact Implicit Scheme

237

iv

FOREWORD

The work reported 1n this program was performed jointly with Avco
Research Laboratory/TEXTRON and Advanced Scientific Computing Ltd. Special
acknowledgement is given to the contributions madia by the following:
P.F. Galpin
B.R. Hutchinson
G.D. Raithby
M.J. Raw

1.0

SUMMARY

.:

The Hybrid-Upwind finite volume discretization schemes adopted in

numerical combustor codes for routine use is plagued with excessive numerical
diffusion errors, which generally preclude accurate quantitative
calculations. In addition, the overall solution algorithm as well as the
resulting discrete algebraic equations require excessive computational
resources for a solution on grids of practical significance, either due to the
inadequacy of the coupling approximations introduced or the unacceptably slow
convergence characteristics of the solvers used. The National Aeronautics and
Space Administration, under the HOt Section Technology program, sponsored
efforts to identify and evaluate quantitatively potentially attractive schemes
that promise improved discretization accuracy, while yielding stable and
physically meaningful solutions. Also considered in this program were various
means of enhancing convergence thus reducing the computational costs
associated with a solution, especially for three-dimensional applications.
This report describes the details of one ,such study.
The desirable attributes of optimum discretization schemes and convergence enhancement techniques adopted in this study for Initial qualitative
assessment, concern issues of accuracy, stability (robustness), efficiency,
storage requirements and ease of implementation in a, three-dimensional code
structured in TEACH methodology. The initial evaluation of more than ten
potential techniques considered was primarily based on examination of accuracy
and linear stability of the resulting difference equations via evaluation of
the properties of the coefficient matrix, Taylor series analysis and existing
heuristic stability analyses for iterative solvers commonly used in segregated
solution procedures, as well as criteria for implementatlonal details. This
effort was aided by information available in the literature, prior experience,
qualitative/quantitative assessment derived from a deep appreciation of both
the schemes and requirements of practical engineering computations. In addition a Technical Advisory Committee with a broad base of 1n depth experience
contributed significantly.

-1-

Upon completion of this initial evaluation, four of the most promising

techniques were incorporated in a variant of 2D-TEACH code for further quantitative evaluation. The particular schemes selected address issues of both
discretization accuracy and convergence enhancement and include:
i)
Second order upwind differencing scheme
it)
Variants of skewed upstream differencing schemes
lit).
Variants of compact implicit method
iv)
Strongly implicit procedure accelerated by:
a) Conjugate gradient algorithm
b) Block correction technique
c) Additive correction multigrid algorithm
Several two dimensional test problems, for which analytical or "exact"
numerical solutions exist or can be generated with relative ease, were adopted
to quantitatively evaluate the performance of the selected techniques regarding issues of accuracy, stability and nature of solutions, and the associated
computational cost. The test problems included general scalar transport and
various flow problems (both laminar and turbulent) and were specifically
designed to evaluate the sensitivity of the above techniques in response to
convection, diffusion and source terms of a general conservation equation, as
well as the nature of the coupling approximations introduced in the particular
incompressible flow solver. Such two-dimensional exercises identified the
variants of skewed upstream differencing schemes, coupled with the strongly
implicit procedure accelerated by the additive correction multigrid algorithm
as deserving further evaluation in three dimensions. The selection criteria
were based on the practical concerns of best accuracy oh coarse grids, robustness and stability and efficiency. Subsequently, these schemes were incorporated in a variant of 3D-TEACH code and were further evaluated regarding
issues of accuracy and convergence enhancement characteristics in modelling of
a jet in cross flow.
This study has clearly demonstrated that appropriate solution techniques
for incompressible, turbulent, viscous, recirculating flows in current use
benefit substantially from the introduction of the convergence enhancement
techniques adopted here (up to 40 percent). Furthermore, the various improved
discretization schemes considered, specifically variants of skewed upstream
differencing schemes with physical advection correction, usually provide a

-2-

significant improvement in accuracy relative to Hybrid differencing. The CPU

times to obtain a solution using these improved schemes are larger than Hybrid
for a given grid but, compared on the basis of equal accuracy, CPU improvements ranging from a factor of 2 to a factor greater than 50 can be realized
via their use. However, as the results of this study are based on a limited
and carefully selected set of problems, the performance of such schemes should
be examined on a wider range of problems and as such this study has presented
only a phase of progress in the complex, but steadily evolving subject concerned with the development of improved discretization techniques and effective solution algorithms. A number of questions have been generated as a
result of this study and it is expected that answers to these questions will
lead to further improvements in accuracy and computational efficiency.
Finally, it is noted that the enhanced solution and discretization
schemes considered represent but a portion of new technology that would be
optimally utilized if freed of the constraints improved by TEACH, which is
structured to suit different methods With different priorities.

-3-

2.0
2.1

INTRODUCTION

BACKGROUND

The trend in aircraft gas turbines continues to. maintain or increase

engine life while improving performance by going to higher operating temperatures and pressures. This imposes increasingly stringent requirements on combustor designs. Heat release per unit volume becomes higher, the uniformity
of combustor exit temperature becomes more critical and the liner environment
is more hostile.
Hot section components such as the combustor liner and turbine airfoils
account for a major portion of engine maintenance costs. A predominant failure mode for current combustor designs is creep/low cycle fatigue interaction
which leads to cracking. Critical liner loads are caused by cyclically
imposed thermal gradients. These gradients are generated by the local radiative and convective heat fluxes within the combustor.
The uniformity of the combustor exit temperature profile is very important to cycle and turbine design. The structural design of a turbine is a very
complex process. Stress limits are a function of overall engine operating
conditions, material characteristics, geometry and local temperatures. Therefore, spatial variations in the temperature profile entering the turbine will
produce local regions of high temperature which usually determines this limiting design condition. The turbine generally will be designed to withstand
this local high temperature at all locations, if the profile is known (either
through analysis or previous development testing). Since the rest of the turbine is exposed to lower temperatures, the engine will be forced to operate at
an average temperature which is lower than it would withstand if the profile
were more uniform. In effect, the material is stressed to the limit of its
capabilities only in the region of local high temperature. The reduction in
average cycle temperature leads to a corresponding reduction in overall engine
performance. In addition, local high temperature regions can lead to premature failure when they have not been identified during development. For these
reasons a comprehensive combustor model, which can accurately predict performance and flow field characteristics, is particularly valuable to engine
designers.
-4-

Gas turbine combustion involves extremely complex physio-chemical processes, including three-dimensional two-phase flow dynamics,1'turbulent mixing,
fuel evaporation, radiative and convective heat transfer, and chemical kinetics. Conceptually, distinct zones can be identified within a convection*!
combustor. These characteristic regions are dominated by one or more of these
processes. In the primary zone, evaporation and mixing rates are comparatively high, and flame stabilization is achieved by recirculation of partially
burned products. In the secondary and dilution zones, partially converted
products are allowed to react further, and additional air is added to reduce
the final gas temperature to a level and pattern acceptable to the turbine.
An effective combustor aerothermal model has to treat each of the individual processes, as well as their nonlinear interdependences, in a cost
effective manner. The end goal is to develop a design and analysis tool which
is significantly less expensive to employ than actual testing of alternative
hardware designs. It must be capable of determining the effects of geometry,
fuel flow, and inlet airflow conditions on the following global design parameters:
i
i) Combustion efficiency
:
ii) Total pressure loss
iii) Exit temperature distribution
iv) Ignition, stability, and relight
v) Pollutant levels
vi) Thermal loading at the walls
vii) Characteristic times
Currently available, three-dimensional, fully elliptic combustor models
frequently employ numerical algorithms and discretization procedures derived
from the original work performed at Imperial College as embodied in the TEACH
family of codes.
Specifically, these solution methodologies, coupled
with appropriate physical models to describe the various physical and chemical
processes, generally employ a segregated method and a staggered computational
grid to solve the algebraic transport equations, derived from the parent differential equations using a finite volume method. The usual formulation is
intended for fully incompressible or subsonic flows. Segregated methods solve
the algebraic equations for each variable separately.

-5-

According to the finite volume procedure, fluxes by convection and

diffusion are required at control volume faces. The diffusive fluxes are
obtained by assuming a linear variation of variable between nodes. In the
original discretization scheme, now commonly, used for routine engineering computations notwithstanding its inherent low accuracy, the value of a convected
variable was approximated by a Hybrid scheme that combines Central and Upwind
differencing.
.
TEACH/derivative methods have proven to be robust, and the results
obtained for a given mesh are generally user independent if convergence isachieved. The number of iterations to reach convergence is small for coarse
meshes but escalates quickly as the mesh is refined. The solutions of momentum equations in each cycle are inexpensive compared to the solution of a
Poisson-like equation for pressure; improving solution economy, therefore,
requires reduction in the cost of the pressure equation.
The Hybrid approximation of convected quantities that results in a
robust method, introduces significant numerical error into the solution. One
of the manifestations of the error is that sharply varying features of the
flow are smeared in the solution.
The need to alleviate the above shortcomings related specifically to
computational and algorithm dependent aspects of comprehensive numerical combustor models in current use, were clearly identified (re-iterated) along with
improvements required in physical models for various phenomena during Phase I
/1 O A \
of the Aerothermal Modelling Program sponsored by NASA '
. The three
participants performed independent assessments of the state of the art regarding numerical simulations for combustor performance and concluded that current
models can only qualitatively predict the complex aerodynamic flow field in
combustors. Quantitative characterization of combustor flow fields was shown
to require a significant reduction in the numerical diffusion levels introduced by the current convective differencing practices incorporated in such
models. Furthermore, especially for three-dimensional applications, the computational cost associated with obtaining converged solutions was found to be
prohibitive, thus a clear need to perform economical computations using improved solution methodologies was identified. Improvements in the physical
modellingi areas related to the description of; fuel spray distribution,
details of multi-phase flow, chemical kinetics, radiation, heat transfer, etc.

-6-

were structured to follow the development of an accurate, cost effective,

aerodynamic model.
.
Another problem Identified during Phase I was a deficiency in test data
from well defined and documented benchmark experiments which could be conveniently used to verify the overall code and its constituent models.
The smearing of sharply varying features of flow (fronts) associated
with the numerical diffusion of low order convective differencing schemes can,
in principle, be overcome by the use of several alternate practices. .The simplest to implement amongst such potential techniques concerns the use of mesh
refinement. However, for three-dimensional applications especially, the
greatly increased number of grid points and therefore iterations required to
reach an accurate converged solution, make the cost of this approach prohibitive. NASA has therefore, evaluated other practices using alternative differencing schemes, less susceptible to numerical diffusion, under a separate
program.
The evaluation included Quadratic Upstream Interpolation for
Convective Kinematics (QUICK) and variants of accurate Skewed Upwind Differencing Schemes (SUDS) appropriately blended with less accurate, but stable
Upwind differencing. These approaches improve accuracy at the expense of a
significant increase in the number of iterations required for a converged
solution. Therefore, two alternative solution algorithms primarily designed
to enhance convergence, SIMPLER and the Pressure Implicit Split Operator
(PISO),
were also evaluated. The general conclusion of this study was
that no single scheme emerged displaying superior performance for all flow
situations.
2.2 OBJECTIVES AND APPROACH

The overall objective of the present study is to investigate methods of

improving the accuracy and efficiency of numerical techniques used to predict
incompressible, turbulent, viscous, recirculating fluid flows. However, accuracy issues related to improved modelling of turbulence are not explicitly
considered; emphasis is solely on reducing the discretization error and
solution cost via improved solution algorithms and differencing schemes.
Furthermore, in this effort accuracy is emphasized over cost when the
requirements to reduce numerical diffusion and solution cost are incompatible.

-7-

The approach adopted here to accomplish the above stated objectives was
a structured effort to identify from the literature and/or other sources over
ten potentially promising techniques to be subsequently assessed for discretization accuracy, solution stability and overall algorithm cost effectiveness.
Included in the selection were convergence enhancement techniques to Improve
the cost effectiveness of the solution algorithm, as well as those including
more accurate discretization schemes. The Initial assessment was primarily
based on examination of the accuracy and linear stability of the resulting
difference equations via evaluation of the properties of the coefficient
matrix, Taylor series analyses and existing heuristic stability analyses for
iterative solvers of segregated solution algorithms adopted here. Cost effectiveness was judged on the combined outcome of the foregoing assessment.
This quantitative/qualitative initial evaluation yielded the four most
promising techniques compatible with the objectives, which were subsequently
incorporated in a variant of 2D-TEACH code for further quantitative evaluation. The stability, accuracy and cost effectiveness of these techniques were
then examined by computing a number of scalar transport as well as various
laminar and turbulent flow test cases. These test cases, for which analytical
or "exact" numerical solutions exist or can be generated with relative case,
display predominant features encountered in gas turbine combustors, i.e., the
delicate local balance between the influences of convection, diffusion and
sources of a general transport equation.
Such two-dimensional exercises identified the appropriate techniques to
improve solution accuracy and overall cost effectiveness in three-dimensional
applications via convergence enhancement and accurate discretization. These
techniques were then incorporated in a variant of 3D-TEACH code and subsequently their performances were assessed in a test case of modelling a row of
jets in a cross flow by comparison with experimental data and prior
computations.
2.3

ORGANIZATION

In Section 3 a brief discussion is provided to examine the computational

details of the solution algorithms in current use for the class of flow problems considered in this study. Sections 4 and 5 present a detailed account of
the schemes. In addition, details of the particular assessment procedure

-8-

adopted, regarding issues of solution cost effectiveness by consideration primarily of accuracy and stability (discretization as well as iterative solvers)
aspects, prior to quantitatively evaluating the schemes in a variant of
2D-TEACH code, is examined. Section 6 provides detailed derivations of the
selected schemes, while Section 7 describes the results of the corresponding
computations obtained for two-dimensional test cases. A similar discussion is
presented in Section 8 for three-dimensional applications. Finally, in
Section 9 concluding remarks are given and recommendations for future work are
outlined.

-9-

3.0 OVERVIEW OF THE COMPUTATIONAL DETAILS OF THE SOLUTION

ALGORITHMS FOR FLOW PROBLEMS
3.1

INTRODUCTION

The majority of heat transfer and fluid flow problems of engineering

interest are analytically intractable. Approximate solutions of varying
degrees of sophistication are currently obtained via numerical methods^ Combustion research, in particular, is presently benefitting from the enormous
potential, versatility and reliability offered by these numerical procedures.
As combustor developments move into the widely varying, challenging, dynamic
field of innovative and conventional power generation including gas turbine
engines, the designer's task can be greatly aided by prior prediction via a
mathematical model. Furthermore, the current interest in fuel efficient, low
maintenance, clean engines has brought forward the neccessary to gain refined
insight into the innumerous problems. The cut-and-try design methods of the
past which cannot answer the above questions in a satisfactory detailed manner
are being guided, if not replaced, by the superior predicting capabilities of
the present approaches.
A common approach adopted by the above numerical predictive techniques
for analysis of viscous recirculating flows, such as those arising in gas turbines, involves the formation and solution of discrete algebraic equations
that represent the conservation of mass momentum and other relevant variables. Various techniques ranging from simple Taylor series expansion to
finite volume and finite element methods are currently used to affect such a
transformation that systematically reduces the governing partial differential
equations to an algebraic set. In the finite volume technique, adopted as the
primary scheme in this study, the algebraic set is generated by integration of
the governing partial differential equation over spatial dimensions spanned by
discrete (finite) control volumes.
For each control volume of the relevant variable there exists an algebraic equation to be solved. Assembly of such nodal equations with due
account for nonlinearity and intervariable coupling yields the equation set

-10-

that is solved iteratively upon specification of necessary initial arid boundary conditions. As the algebraic nodal equations consistently approximate
the partial differential equations (valid for control volumes of infinitesimal
size) over finite spatial dimensions, increased numerical accuracy in the finite volume solution is usually achieved through mesh refinement. However, for
a given problem domain, decreasing the control volume sizes increases the number of control volumes, thus increasing the size of the discrete equation set
to be solved and hence presenting a greater computational task. The practical
limitations of mesh refinement are severe; prohibitive demands may be placed
on computing resources before a grid independent solution is attained. This
problem is particularly acute for many three-dimensional fluid flow and heat
transfer problems.
The need for greater computing economy in predictive techniques has
opened up two basic avenues of research. The first involves the development
of more efficient solution methods designed to reduce the computational
resources required to affect the solution of the algebraic equations resulting
from the finite volume method or otherwise. The second involves the development of more accurate discretization schemes, aimed at reducing the number of
control volumes (nodes) needed to achieve a given accuracy in the finite volume solution. Closely tied with this issue are considerations regarding false
diffusion or spurious spatial conditions. Effectively dealing with the above
problems of solution accuracy and efficiency forms the main themes of this
study and such efforts should contribute to reducing both computing time and
storage requirements for engineering predictions. The latter is essential if
the scope of tractable three-dimensional problems is to be expanded.
In the following sections are presented the basic building blocks appropriate for the development of such techniques
3.2 THE MATHEMATICAL PROBLEM

The conservation laws governing heat transfer, fluid flow and other related processes are generally expressed in terms of integro/differential equations derived via continuum or microscopic considerations. The identification
of all the relevant phenomena that can be described by a representative equation of a common form is the first step toward constructing a general solution

-11-

procedure. Specification of two additional algebraic relations completely

defines the mathematical problem. The first of these describes the thermodynamic and transport properties of the fluid, e.g., the equation of state
relating local values of density, pressure and temperature. In the latter
category are the boundary conditions that uniquely specify the problem. These
usually take the form of either specification of the value of the dependent
variable at the boundary, or the value of the associated flux or a combination
of the two.
3.2.1

General Conservation Equations

The differential transport equations listed below apply equally to both

laminar and turbulent flows (the latter referring to instantaneous description
of the flow). In Cartesian tensor notation, the equations for mass, momentum,
energy and species conservation are:
Conservation of Mass

Conservation of Momentum

1fj

_s

=Q

(32)

ax

Conservation of Energy

. a

(pu.h
j -

Conservation of Chemical Species

. ^ a_ (pu.m. - J

.)

, .v
'

(3 4)

In the above the x are the three Cartesian coordinates, the u. are the
three components of the velocity vector; p is the pressure; p is density.
S

, S. and S . are volumetric rates of body forces, energy generation

n
mi

-12-

and chemical species generation respectively.' h denotes the specific stagna

tion enthalpy defined b y :
, . - - .
1 2
h = Cp^ mjT + U1

.-..;,..

(3 5)

'

where c
is constant pressure specific heat capacity, m is mass frac J
'
tion of chemical species i, T is temperature. In equation (3.2) t. . is
1 j
the stress tensor with components defined in a general form by:

au

au

au

a
. . = Kronecker delta
i J
v is laminar viscosity. In the energy and species equations the J
and
Jm . terms refer to diffusion fluxes and can be written as, if expressed by
t fJ
Fourier's and Pick's law of diffusion:

(3
M

am.

(3 8)

'

where a.h and a m. are respectively the Prandtl and Schmidt numbers,
i
3.2.2 Modelled Form of the Conservation Equations
For the description of turbulent flows the above Instantaneous conservation equations are transformed Into ensemble-averaged equations by an averaging operator that assumes rapid and random fluctuations about an ensembleaveraged value. The ensemble-averaged equations are similar to those for the
instantaneous equations except for the appearance of the additional terms containing correlations of fluctuating components, Reynolds stresses, turbulent
diffusion fluxes, etc. In order that these equations form a closed set, these
additional terms must be related to known or easily calculated quantities.

-13-

Following the generally adopted practice of Launder and Spalding,

the additional terms are modelled as diffusion processes 1n order to make the
ensemble-averaged equations compatible with those for laminar flow, I.e., the
additional terms are expressed In terms of gradients of average quantities as:

,aui

(3.11)

where k is the turbulent kinetic energy ( = u u,/2); y 1s a turbulent

viscosity; the superscript ' stands for fluctuating quantities; u", "R and
m represent ensemble-averaged values of velocity, stagnation enthalpy and
mass fraction of the chemical species respectively; a.h,t^ and am-,t^ are
turbulent Prandtl and Schmidt numbers respectively and are usually assigned
constant values determined by experiments.
Finally, the modelled forms of the ensemble-averaged conservation equa
tions are given as, dropping the bars over time-averaged quantities:

j - S p-

(3.13)

Iat

~
-14-

3m.

l
where S , S , S. and S 'comprise terms arising from sources of nonp
n>j n

uniformities in fluid properties.

To close the solution of the above equation set, spatial and temporal
evolution of y has to be described. This is generally achieved by developing two transport equations for the turbulent kinetic energy, k, and its dissipation rate e, = (v/p) 3ul/3x.) , an approach moderate in complexity,
but sufficient for most engineering flows (8). The ensemble-averaged equation
for k and c are given by:
Turbulent Kinetic Energy Equation

)-Sk-Grpc =0

(3.16)

Dissipation Rate Equation

u e
s
f?
e/k = 0
at (P^ + IT"
ax. (P ji - IT
o If
ax. ) - c - (C,G-C,pe)
\ i

(3.17)

/g\

where C,, C_, o. and a are empherical constants

and S. and
Sc are additional source terms.
From dimensional analysts, the relation between v and these two
variables is given as:
2
wt = P C k /c

(3.18)

C is generally a constant.
3.3 ALGEBRAIC REPRESENTATION OF CONSERVATION EQUATIONS

Review of the above differential transport equations (3.12) to (3.17)

reveals that they are similar in structure and all the dependent variables are

-15-

conserved, intensive properties. If a typical representative variable is

denoted by 0, the general differential equation

ft
where the expressions for r and S can be deduced from the parent equa0
0
'. . .
tions, forms the central theme of discretization studies. The four terms in
the above equation describe successively unsteady, convection, diffusion and
generation/dissipation effects. (In fact all terms not explicitly accounted
for in the first three terms are included in the "catch-all source" term.)
3.3.1

Domain Discretization

In the following sections the finite volume technique is adopted as the

primary means of transforming the differential transport equation into a consistent algebraic set. Furthermore, the discussion will be restricted to twodimensional, steady, incompressible form of the equations, where the density
approaches a constant value. Thus, the temporal derivative of density in mass
conservation vanishes with corresponding modifications in the remaining equations.
Generation of the nodal algebraic relations proceeds by discretizing the
calculation domain in some fashion. To prevent the occurences of decoupled
pressure solutions and spatial oscillations of the velocity solution that can
arise in the incompressible limit,
the staggered grid, figure (3.1) first
(9)
used by Harlow and Welch,
is utilized. Recently there have been
approaches that use co-located storage for variables and overcome the pressure
(10)
decoupling problem in alternate ways;
however, their use so far has been
limited. The calculation domain is divided uniformly into non-overlapping and
contiguous control volumes over which mass, energy and species conservation
are to apply. Locating the pressure, temperature and other scalar variables
in the center of these control volumes, the staggered velocity nodes are
located at the faces of the scalar (e.g. continuity) control volume. The control volumes for momentum conservation are arranged so that pressure nodes lie
on the faces of their respective momentum conservation volumes.

-16-

I- V Control' Volume

frjfjTjjj

iff

-1J

Figure 3.1

1 ^

T
p= c

U
-

Control Volume

$ -

~*~-, *

ft

T-

^ 0 Control Volume

Grid Structure Showing Variable Locations

1NW (i - 1, j + 1)

N U

Vp

-f
^H-

tNE (1 + 1,1 1)

.i!
n-7~ J.
yt l
H
-^ ~

8yj

-1

Figure 3.2

1-1 M--f-

Illustration of Grid, Finite Volumes and Notation for a Typical

Finite Volume

-17-

The notation used to denote relative control volume locations is illustrated in figure (3.2). The subscripts E, W, N and S denote quantities
associated with control volumes to the right, left, above and below of an
arbitrary location in the domain denoted by a subscript P. The P nodes of x
and y components of velocity are associated with the pressure nodes directly
to the left and below P.
3.3.2

Discretization of Differential Equations

The differential equation is formally integrated over individual finite

volumes, and the resulting volume Integrals are transformed to their surface
counterparts where possible. The remaining integrals are subsequently approximated by algebraic expressions. In this process, the surface integrals are
Initially subdivided Into piecewise segments, associated with specific grid
points, in a symmetrical fashion. The resulting integrals are then approximated with the aid of the mean value theroem. Finally, the approximated
piecewise surface integrals are transformed Into algebraic expressions connecting the given grid point values. This transformation may be done in various ways, e.g. profile assumptions, finite differencing or local analytic
solutions etc. It should be stressed here that in all cases particular attention must be paid to satisfying certain constraints discussed in
Section (3.3.3). Finally, the discretised equation is assembled as the
approximated integral equation.
To clarify the above approach and to develop adequate appreciation for
the subtleties of the advanced discretization schemes discussed in Section 5,
the following example Illustrates the application of the technique for twodimensional transport of a scalar. This model equation plays a central role
in the development and evaluation of improved discretization schemes.
Equation (3.19) for a general scalar 1n Cartesian coordinates is
expressed as:

fx" <"ui> + fy" <"v> = a7 (r0 fj> * fy" (r0 !J> + S0

where u and v are time averaged components of the velocity 1n the x and y
directions, r0 is the appropriate diffusive exchange coefficient and S0
is the corresponding source term. Integrating the above equation over the
control volume of dimensions ax and Ay, figure (3.2).

-18-

AX Ay
(3.20)

the following form is obtained upon application of the Gauss' theorem:

Ay

AX

Ay
(3.21)

where S0 is assumed to be constant over the control volume; and

the subscripts
'
e, w, n and s represent evaluations along the corresponding control.volume
faces.
To derive algebraic representations of the balance given by
,
equation (3.21), evaluations of the convective and diffusive fluxes, expressed in terms of the nodal values of 0, are required along each of the
four control volume faces in the manner described below.
Diffusive Fluxes

The diffusive flux across the e face, e.g. J , is evaluated as

follows:

JJl.

"V V1

(3.22)

Ay
Implicit in the above expression are the following assumptions regarding variable profiles along the e face:
i) r 0 is constant
ii) 80/ax varies at most linearly along the face
iii) 30/ax can be represented by (0 -0 )/AX
Similar relations can be derived for the diffusive fluxes across each of
the other faces.

-19-

Convectlve Fluxes
For the convective fluxes an approach similar to that adopted for the
diffusive fluxes can be used. In general, this approach will lead to any number of discretization schemes including the five point-operator discretization
schemes, such as Central differencing and Hybrid differencing
or the
nine point discretization schemes such as the Skewed Upwind differencing
scheme of Raithby.(12) In general, these schemes result 1n discretizations
which suffer from false diffusion or spurious spatial oscillations. In both
instances, these errors can be shown to arise from an inadequate evaluation of
convective fluxes across the control volume faces.
3.3.3

Discussion on the Discretization and Solution Schemes

As there exist a myriad of arbitrary ways to reduce the differential

transport equations to their corresponding algebraic sets, some guidelines
have to be followed to arrive at rational;choices. First, a consistent discretization scheme should reproduce the exact differential solution in the
limit of an Infinite number of grid points. Section (4.2) discusses the properties likely to be displayed by exact solutions. A practical appreciation
of the above leads to the notion of the behavior of solution accuracy with
modest grid refinement achievable, especially for three-dimensional problems.
A closely related issue regarding order of accuracy is to avoid the false diffusion of first order accurate schemes or the instability of conventional
second order accurate differencing for convection.
Thus, some descriptive attributes of an "ideal" discretization scheme
might be stated as follows:
i) Generality and ease of application.
A discretization scheme should be capable of treating many simultaneous
and strongly coupled phenomena in practical configurations of engineering
geometry, i.e., a wide range of flow regimes in the presence of turbulence,
heat and mass transfer, chemical reactions in arbitrary geometries. Furthermore, ease of application is also an important consideration. Generally
speaking, the more complex the scheme, the more troublesome the imposition of
boundary conditions.

-20-

i i) Accuracy and Economy

,,

., .

Minimization of solution error and computational cost are increasingly

assuming significance as the primary motivation for the-development and
assessment of improved discretization and solution schemes. Taylor Series
Expansion error analysis favored heavily in the past in the delineation of
order of accuracy, has important restrictions on its validity in certain1 circumstances. However, one of the advantages still offered by such analyses is
an indication regarding the nature of the error, i.e. dispersive'('unbounded)
or diffusive (dissipative). It also indicates the likely rate of reduction, of
approximation error with mesh spacing as the latter, in principle, becomes
vanishingly small.
Currently other techniques including comparison with analytical solutions of idealized model problems are being used to guide the construction of
discretization schemes. This is notwithstanding the fact that such model
problems cannot embody all the pertinent characteristics of practical flow
problems for which these schemes are being developed.
Economy of solution for schemes specifically refers to the relative cost
of obtaining solutions of specified accuracy to a given problem by different
schemes. This is generally influenced by: the accuracy of the schemes; the
number of arithmetic operations required in coefficient assembly; and the type
of solution algorithms which the discretized equations admit.
iii)

Transparency and Reliability

As the discretized algebraic equations are the finite volume (or otherwise) analogs of the parent differential transport equations, the solution
behavior displayed by them should reflect the fundamental transport physics,
embodied in the latter, regardless of the solution variables. This is discussed further in Section (4.2). Furthermore, appreciation of the simplifying
assumptions and approximations embodied by discretization schemes should.help;
to understand and monitor their behavior.
3.4 GENERAL SOLUTION METHODOLOGIES FOR INCOMPRESSIBLE FLOW PROBLEMS

The conservation equations of mass, momentum, energy and species presented in Section (3.2.2), equations (3.12) to (3.15), are both nonlinear and
coupled. After suitably discretizing them to generate the corresponding algebraic set, it is convenient to linearize them to fully utilize the advantages

-21-

offered by the well developed techniques for the solution of linear algebraic
equations. Iteration is then required to account for the nonlinearities.
Generally, there are two classes of methods used to solve the linearized
algebraic equations for pressure and velocity for incompressible flows; direct
methods
and iterative methods.
Direct methods are usually not
used because of the excessive computational resources that are needed. However, because of the implicit nature of pressure and the pressure ~ velocity
coupling the development of iterative methods can be difficult.
3.4.1

Linearization of Mass and Momentum Conservation

For two-dimensional incompressible flows where the density is a constant, the conservation of x and y momentum yield finite volume equations of
the following form for u and v respectively, figure (3.1).

3 U

P P = VV*
NP

rAy(pE-pP)

NPVNP+bP"Ax(pN'PP)

(3 23)

'

(3 24)

'

NP

In equations (3.23) and (3.24) the summations may imply a five point or
nine point computational molecule depending on the discretization scheme.
From these discrete momentum equations it is seen that pressure differences
appear as separate source terms. These pressure differences arise from the
integration of the pressure gradient volumetric source strengths in the partial differential momentum equations. They are distinguished from the remainu
v
ing source terms bp and b. in order to make explicit the coupling between
pressure and velocity^ The bp and bp terms may contain transient and body
force effects.
The conservation of mass for the p control volume is expressed by:
pAy(up - uw) + pAx(vp- v$) = 0

where a constant fluid density p has been assumed.

-22-

(3.25)

A primitive variable solution procedure must arrive at u,v and p fields

which satisfy equations (3.23), (3.24) and (3.25) for every u , v and
p control volume (respectively) in the solution domain.
From Section (6.1) it is easy to appreciate that the x-momentum coefficients (a"K and a")
NK and the y momentum coefficients (arD and
v
a...) must depend strongly on the mass fluxes at the faces of the respective velocity control volumes. To illustrate how these mass fluxes are
estimated, consider the mass flux through the west face of the u control
volume in figure (3.2).
U

= pAy(

P + UW
2 '

(3 26)

'

Similarly, the mass flux through the south face of the v control volume in
figure (3.2) is approximated by:
r
= PAx( V

+V

S*)

(3.27)

Mesh uniformity has been assumed in equations (3.26) and (3.27) in that the
velocities at these faces are estimated as the arithmetic mean of the nodal
velocities on either side of the faces. The mass flux through the north face
of the u_ control volume is comprised of two velocities, each occupying half
of the face:

U
AX
AX
n = p~2VP * p ~2VE

(3 28)

'

Similarly for the east face of the v control volume:

e=

^2UP * p^2 UN

(3>29)

From these examples it is clear how the mass flux at any face of a
velocity control volume may be calculated.

-23-

Since the momentum coefficients in (3.23) and (3.24) depend on mass

fluxes, such as (3.26) through (3.29), which themselves depend on the, unknown
velocity field, it is evident that equations (3.23) through (3.25) constitute
a nonlinear equation set. The linearization of equations (3.23) and (3.24) is
accomplished by simply basing the momentum coefficients on mass fluxes that
are calculated from best available velocity fields. Once velocity and pressure fields are obtained which comply with equations (3.23), (3.24) and
(3.25), they become the best available field for the next coefficient iteration. The coefficient iteration is terminated when the velocity and pressure
fields converge.
3.4.2 Solution of the Unreduced System of Equations
Linearized equations (3.23) through (3.25) can be expressed in the corresponding matrix form as:

{i H M

(3.30)
(3.31)

01 ]

A*1

{*}

{v} -

{0}

(3.32)

or equivalently.
(3.33)

-{>*}

where

A'

AU

III
-24-

u-

The solution of equation (3.33) can be obtained directly by inversion of

[A ], by Gaussian elimination or their variants. However, the computational
effort and storage requirements of these methods often are prohibitive. The
computational effort requirements are further increased by having to solve
(3.33) for each coefficient iteration.
To reduce the computational resources needed to solve equation (3.33)
iterative solution methods are often considered. However, examination of
[A ] reveals that there are entries along the principal diagonal which are
zero. These zero entries, which arise because pressure does not appear in the
equation representing mass conservation, makes the development of iterative
methods particularly difficult. Iterative methods must ensure that the implicit influence of pressure on velocity through momentum conservation is
accounted for appropriately. Iterative methods based on equivalent set of
reduced equations overcome this difficulty.
X

;
Solution via a Generalized Segregated Approach
Using the Direct Simultaneous Variable Solution (13) a sound basis can
be established for the development of a generalized segregated approach. The
velocities appearing in equations (3.30) through (3.32) can be expressed in
terms of the pressure by multiplying equations (3.30) and (3.31) by [A ]
v -1
and [A ] respectively.

3.4.3

= [A U ]

= [AY1

bu

- [DJJ] p

(3.34).

bv

- [Dj]

(3.35)

where

[Dp] = [AV]~Vp]

and where [0 ] and [D ] represent the influence of all pressures in the

calculation domain on the u and v velocities respectively. Substituting
for the velocities from equations (3.34) and (3.35) into (3.32) there results

-25-

[A P ]

|P}

= {b P }

'-'' (3.36)

where
[A P ] =

|bP}

- [MU] [DjJ] - [MV]

- [MU] [AY1

[Dp]

bU - [MV] [AY1

Such an approach still requires excessive storage, and effort (in general, less than those for the unreduced system discussed in the previous
section) to invert and store [AP], [D ], [0 ] etc. Van Doormaal
through a detailed study regarding the nature of [o"] matrix established
how the mechanisms of convection and diffusion can spread the influence of
pressure throughout the complication domain and analyzed the implications of
adopting localized approximations (a feature that is prevalent in most segreu
v
gated approachs) for it. Denoting by [D ] and [D ] the approximate
evaluations of [D^] and [0^], a generalized segregated approach that
employs iteration for these approximations is now formulated. Taking the
value of pressure at the k'th segrated iteration to be |p*L the estimate of
pressure, the corresponding {u*} and |v*| velocities which satisfy
equations (3.30) and (3.30) are given by:

(3.37)

(3.38)

Because jp*} is not correct, then |u*[ and |v*| are not correct. To
obtain improved estimates of velocity given by J1T| and {7}, it is necessary to
subtract out the effect of |p*| on |u*l and |v*| and add in the effect of an
improved pressure estimate given by |"p{. This can be accomplished in an
approximate way by subtracting -[~DU] |p*| from and adding - [DU] |"p}
to equation (3.37), and subtracting - [IT'] | p*| from and adding
- [D] {?[ to equation (3.38). As a result;

(3 39)

(3.40)

-26-

where

Requiring that the velocities /Hi and < v> satisfy mass conservation
[Mu] {u} + [Mv] |v|= |o|

(3.41)

and substituting for velocities from equations (3.30) and (3.40) the following
equation for pressure results,

(3.42)
where
[A p ] = - [Mu] [Dp1] - [Mv] [Dp7]

The success of the generalized segregated approach described by

u
v
equations (3.37) to (3.42) depends on how accurately [0 ] and [DQ]
are approximated and how readily [0 ] and [01 are evaluated and the
solution for pressure is determined. To ensure that the solution for pressure
is easily found it is convenient to relate each of the velocities in terms of
nodal values of pressure and velocity (cf. equations 3.23 and 3.24).
*ll

^^V

up

= fjj, - du (pE - pp)

vp

= ^

-27-

- d7 (pN -

Pp)

(3.43)

(3.44)

where

x\
* -u * *
up = up * d (pE - pp)
/v
V
p

*
-\f *
*
d (P
P
p *
N - P>

and where for SIMPLE/1* e.g.

= ay/ap

. V
dTV = Ax/a
p

Substituting for the velocities from equations (3.43) and (3.44) into
(3.25) the equation for pressure is given by, figure (3.2):

3P

PP

(aPp)

NP+ bP

NP

where
NP

p
E

a[J

-u
= PAy dp

-V

= pax dp

-* +
-28-

pa^w -

(3 45)

'

4.0 ASSESSMENT OF IMPROVED DISCRETIZATION SCHEMES AND SOLUTION ALGORITHMS

FOR GENERAL (INCOMPRESSIBLE) FLOW PROBLEMS
4.1

INTRODUCTION

, .

Recently substantial progress has been made in the development numerical

procedures for the solution of differential equations governing fluid flow, !
heat and mass transfer for engineering applications. However, there remain,
in practice, certain deficiencies in these methods that need to be resolved
before they can be used reliably and efficiently.
Prediction of complex turbulent flow structures including separated
zones using "conventional" discretization and solution schemes runs Into serious accuracy issues, as such schemes display appreciable deterioration in
accuracy due to misalignment of the flow and computational mesh. Most practical schemes in current use artificially dissipate steep gradients due to
limited spatial resolution capabilities. The resulting effect, 1n some cases,
' .
. - i'

may completely overshadow real effects such as physical diffusion due to

turbulence.
Alternative schemes specifically designed to alleviate such dlssipative
errors succeed only partially, as the solution 1s usually plagued with Improperly bounded, non-physical oscillations. Associated with such behavior are
problems of numerical instability due to unfavorable conditioning of the
coefficient matrix.
.
The discretization errors of diffusive schemes can, in principle, be
alleviated through mesh refinement. However, such an approach is not always
feasible nor effective, as the solution cost and resources can become prohibitive especially for three-dimensional problem. What 1s required 1s a stable
scheme that yields bounded solutions while minimizing the Influence of diffusive errors. Rational construction of such schemes through careful analysis
of the properties of exact solutions for the differential transport equations
is one concern of this Section.
The numerical solution of Incompressible flow problems 1s further complicated due to nonlinearlties and couplings Inherent 1n the corresponding
differential transport equations, thus considerable computing resources are

-29-

required to obtain solutions to the algebraic equation set derived from an

appropriate discretization scheme. If a segregated (SIMPLE-l1ke)
approach 1s adopted to solve these algebraic equations, then. 1n many Instances, a major portion t of the computational cost 1s associated with the
solution of the equation for pressure (or Its correction). Minimization of
the computational costs associated with solving the linear pressure equation,
as well as a descriptive structure analysis of general segregated methods are
also discussed in this Section.
4.2

CHARACTERISTICS. OF ACCURATE DISCRETIZATION SCHEMES

Some guidelines regarding the attributes of an "Ideal" discretization

scheme can be formulated on the .requirement of consistency between the differential and discretized equations. However, before such a task 1s undertaken,
the pertinent characteristics likely to be displayed by the exact solutions of
the differential equation are reviewed.
(16)
Properties of Exact Solution
I) Conservation
For any region D with surface S the volume integral of the differential
equation (3.19) 1s expressed using vector notation as:
J>lp u 0 - ^grad 0f(ftdS - / S. dD - 0

'

(4.1)

where*'1s the unit normal surface vector (positive outward)

II) Boundedness
It is straightforward to appreciate the significance of the Maximum
principle for the simple case of source free conservation equations. These
solutions cannot display positive maximum or negative minimum inside the
region D., I.e., the solution must be bounded in D by its extreme values on S
min (0$) <_ 0 _< max (0$)

(4.2)

Similar bounds can be derived for more general cases of different boundary
conditions and/or finite source effects.

-30-

iii) Levelness
If the differential equation contains only derivatives of the dependent variable 0, then the functions 0 and 0 + C (C is an arbitrary constant)
both satisfy the differential equation,
iv) Transportive
In the regions of flow where convection dominates, the solution is
strongly influenced by upstream conditions. The directional behavior of flow
is generally associated with the existence of domain of dependence notion with
each mesh point. .
.
Guidelines for the Construction of Discretization Schemes
The nodal algebraic equation is an integral approximation to the differential equation and hence should preserve all the important properties of
the latter as described above. Such constraints generally furnish specific
requirements to be satisfied in the construction of discretization schemes.
By adopting a "suitable" approximation scheme for the convection terms
of equation (3.21), the resulting nodal algebraic equation is symbolically
represented by:
=

a'
P
K
NP

VNP* b P

(4 3)

'

In the above (using the conventional notation) the subscript NP refers

to the neighbours of P. The quantity a is the coefficient for a and
a
is tne
im
coefficient for a NP (i.e., a_E for 0_,
NP
E a,,
W for 0,,,
W
etc.). Term b denotes the source term for the control volume P that may
contain physical source or sink and/or a transient storage term for unsteady
problems. The quantities a.._/a
n are generally referred to as influence
Nr r
coefficients and should ideally all be non-negative to satisfy the boundedness
requirement of exact solutions.
The existence of negative influence
coefficients anywhere in the finite volume equations is commonly referred to
as the negative coefficient problem. The appearance of negative coefficients
in sufficient numbers and magnitudes in the finite volume equation set may
violate the smoothness of solutions by generating non-physical overshoots
and/or undershoots and also lead to difficulties in iteratively solving the
algebraic set. Closely associated with the latter issue is the mathematical

-31-

notion of diagonal dominance. The matrix of Influence coefficients Is

diagonally dominant if, for every node P:

NP

!NP
a
p

< 1

(4.4)

with strict inequality holding for at least one node. Clearly negative influence coefficients violate this rule. For a constant set of coefficients,
matrix theory guarantees convergence to a 0 solution provided the matrix of
influence coefficients is diagonally dominant. It is important, however to
make two points regarding diagonal dominance. First diagonal dominance is not
a necessary condition for convergence. Second, the matrix of influence coefficients often is not constant in practice since the coefficients change as
other coupled solution variables change in the progress of the Iterative
solution; the linearization process, whereby the Influence coefficient matrix
is updated, is another potential source of instability in the overall solution
scheme.
An additional property of significance for equation (4.3) is that of
adidivity expressed by:

* pAXAy

(4 5)

NP

unsteady term
This condition is assured provided the convecting mass fluxes used to
establish equation (4.3) obey continuity for control volume P. This property
is important, since it permits a linear function of 0 also to obey the finite
volume equations, levelness property of exact solutions.
Regarding the conservation property of exact solutions, the algebraic
equation set should satisfy the Integral conservation relation (4.1) both
locally and globally. To meet the additional global conservation, flux continuity in the discretized equation at cell boundaries must be assured. Only
under such circumstances will Interface fluxes at adjoining volume cancel in
pairs, leaving the exterior surface integral to be balanced by interval
sources/sinks.

-32-

The transportive property of exact solutions requires, in the limit of

strong convection, the solution of the algebraic nodal equation to display the
pertinent "directional" sensitivity, i.e., nodal values lying outside the
domain of dependence/influence of a particular node should not feature 1n the
discretization. Failure to comply with this requirement can give rise to the
negative coefficient problem described above and hence, from the boundedness
principle, unrealistic solutions may result.
A practical primary qualifier of a discretization scheme is its
accuracy. For problems of interest in this study, a pragmatic interpretation
of accuracy refers to obtaining solutions uncorrupted by numerical diffusion
using practical grid sizes and covering a range of solution variables, notably
arbitrary inclination of flow to grid and arbitrary values of Peclet number.
The issues of conservation and boundedness are closely tied to the overall
question of the accuracy of a given discretization scheme, for which however
it is much more difficult to derive, in general, a priori criteria or measures. As discussed in Section (3.3.3) earlier analyses of accuracy
compris/ I Q\
ing mainly Taylor Series Expansions have recently been shown
to have
serious limitations and shortcomings. A practical measure of accuracy freely
adopted in this study relates to the rate of convergence of a discretization
scheme. Such a criterion displays quantitively the sensitivity of the numerical solution to grid refinement in asymptotically approaching the
analytical solution for model problems. (19)
Analysis of additional characteristics including generality and ease of
application as well as economy and efficient are also important considerations
in the development and assessment of discretization schemes. A concise summary outlining the attributes of desired discretization schemes is provided in
Section (4.4).
4.3

COMPUTATIONAL DETAILS OF THE IMPROVED SOLUTION SCHEMES FOR

INCOMPRESSIBLE FLOWS

Majority of the incompressible flow solution algorithms in current use

adopt a particular form of the Generalized Segregated Approach,
Section (3.4.3), to deal with the pressure ~ velocity coupling problem,
although recently there have appeared algorithms designed to solve the relevant equation set directly. (20 '24) The following discussion will thus focus
on the structure of segregated approaches including implementation and operational details.

-33-

The [DU] matrix in Section (3.4.3) was shown to represent the

influence of pressures on velocities and was later approximated by [o"]
to affect an easy solution for the pressure. However, this simple form
[0] where each velocity is linearly related to the local pressure difference, introduced primarily to minimize computational requirements, has two
major shortcomings.
For high Reynolds number flows of concern in this study, using minimal
or no under-relaxation for the solution that implies large effective time
(22)
steps for a time marching scheme,
pressure differences far upstream of a
velocity can significantly influence the velocity, thus a relation between the
velocity and only the local pressure difference is not likely to be appropriate. The implication is that the use of any segregated method which sim-u
plifies the evaluation of [D ] by relating each velocity to only the
local pressure difference for solving convection dominated steady flows,
necessarily implies considerable under-relaxation for the variables.
The second shortcoming of this simple form of [D ] is that, without
taking special care, the convergence of the segregated method could degrade
significantly with grid refinement. To illustrate this, consider using a
segregated method that adopts the simple form of [5"] for a high Reynolds
number problem using two grids, one with relatively few control volumes,
another with a large number of control volumes. Suppose that on the coarse
grid a value for E, defined as a multiple of the maximum of explicit time
step,
is chosen such that the simple form of [Q ] was appropriate.
(Discussion here refers to time integration of equations to reach steady
state, although relaxation and time integration approaches can be related by a
simple relationship that exists between E and the more conventional underrelaxation parameter used in study). Then on the fine grid choose E so that
the effective time step is the same as that for the coarse grid. Now, because
the effective time steps are the same for both grids, the spatial distribution
of the normalized influence of pressure differences for both grids will also
be similar. Therefore, on the coarse grid where the E was chosen to that the
simple form ["5 ] would be appropriate, the majority of pressure difference influence would arise in the distance given by the coarse grid spatial
step size. However, on the fine grid, the same distance is covered by a number of fine grid spatial steps. As a result, on the fine grid the velocity
may be influenced by more than one fine grid pressure difference and the

-34-

u
simple form of [D ] will not be appropriate. Only by reducing the value ,
of E (or the under-relaxation parameter) will the simple form of [Ou] be ..
appropriate for the fine grid. But in the process of reducing the value qf E
(or the effective time step)
the number of coefficient iterations (or ti?ne
>
steps) required .to achieve convergence (or steady state) is increased and,
subsequently, the rate of convergence is reduced. To increase the implicitness of the pressure ~ velocity coupling of segregated .approaches, improved
pressure algorithms notably SIMPLER,(1) PISO,(6) SIMPLEC(15) and
others (14) have been proposed and will be used freely throughout this study.
To provide a background for appreciation of some of the techniques considered in this study, it is also appropriate to discuss the solution of the
pressure equation of the SIMPLE algorithm,
equation (3.45).
The solution of the pressure equation is a major component of the segregated approach and can represent a significant portion of the total cost of
solving a flow problem. It is therefore a high priority to solve for pressure
in an efficient manner. As discussed previously, the use of direct solution
methods tends to be unattractive due to large storage requirements and
computer effort. Extremely fast Poisson solvers are available
but these
are not applicable to the pressure equation. Furthermore, the coefficients of
the pressure equation change each coefficient iteration so that sparse matrix
solvers that are applicable require a new decomposition at each iteration.
Iterative methods such as Successive Over-relaxation and line by line methods
using the Tri-Diagonal Matrix Algorithm,(25) Stone's Strongly Implicit
(26)
Procedure
and the Modified Strongly Implicit Methods of Schneider and
(27)
Zedan
are better suited to this application. Finally, because for incompressible flows, the pressure is determined by solving a symmetric positive
definite matrix equation, equation (3.42), the Chebyshev and Conjugate
(25)
Gradient techniques
may be used to accelerate the convergence of many of
the methods listed above.
With the use of iterative solution, methods for the solution of the
pressure equation, termination of the iterative procedure becomes an important
consideration. If iteration of the pressure equation is terminated before
sufficient convergence is achieved, mass conservation is poorly satisfied by
the corrected velocities. Since it is usual for only one iteration of the
segregated method to be performed, these corrected velocities are then used to
calculate new coefficients of the linearized algebraic equations. Propagating

-35-

the error in this fashion can result in divergence or slow convergence of the
coefficient iteration. On the other hand, it is uneconomical and wasteful to
solve the pressure equation to a tight convergence at each segregated
iteration.
Let ||rp|| represent a measure of the residual of the pressure
equation after each 1 iterations of the iterative method used to solve the
pressure equation. One measure of the residuals is given by the Euclidean
norm where:
1/2

(4.6)
NP

AV

and where^rr'is the summation of all interior volumes. An alternative meaAV

sure is provided by the sum of the absolute values of the control volume
residuals.

a p

,P M

pp

(4.7)

NP

AV

A convenient method of terminating the pressure iteration is given by:

(4.8)

where the 0 superscript on ||r || is used to denote the initial measure of

the residual for 1 = 0. Equation (4.8) guarantees that iteration has reduced
the residual of the pressure equation to at least the fraction y and can be
used for most problems with optimal values of the residual reduction factor
Y typically ranging from 0.25 to 0.05. This avoids costly trials to determine a suitable value of the convergence criterion. Additionally, the number
of iterations on the pressure equation for each coefficient iteration is
roughly the same. (19)

-36-

4.4

SELECTION OF THE ASSESSMENT CRITERIA AND METHODOLOGY

The desired characteristics of a discretization scheme in terms of preserving the basic properties of the exact solutions (i.e. conservation,,
boundedness, levelness and transport!ve) as related to issues of accuracy and
stability of solutions have been discussed in detail in Sections (3.3.3) and
(4.2). This Section will expound on those ideas and introduce additional
auxilliary criteria such that a comprehensive list of attributes is developed
to critically evaluate the schemes considered in this study.
It was established in Section (3.3.3) that a priori, assessment of
accuracy of a scheme in a complex turbulent flow environment of the kind considered in this study was extremely difficult. This was shown to be primarily
due to the inadequacy of the error analysis provided Taylor Series Expansion
for the relatively coarse grids used (especially for three dimensions) as well,
as the limited guide offered by comparison with model solutions. However,
such methods could be effectively used to economically identify schemes of
maximum potential. Also when evaluating the performances of selected schemes
in accurately predicting the engineering details of turbulent flow, means have
to be provided to delineate the uncertainties and error introduced by the
turbulence model.
, ..
Stability of the iterative solvers, used primarily to r minimize computational costs associated with a solution, were also shown to be strongly
influenced by the condition of the matrix of influence coefficients. The
impact of such restrictive requirements can, in principle, be alleviated or
lessened by solvers that are more transparent to the details of the discretization scheme (stable). However, development and assessment of such solvers
to cover a wide range of condition numbers, as variables continually change in
the course of the iterative solution, is a very demanding undertaking.
The specific definition of economy or alternatively cost effectiveness
of solutions, as adopted in this study, pertains to the relative costs of obtaining solutions of specified accuracy to a given problem by different
schemes. The cost will generally be a complex function of the accuracy of the
schemes, the nature and the number of computer operations required in coefficient generation and assembly, the type of solution algorithms which the
discretized equations admit, and the details of the computer architecture used

-37-

for solutions. Economic considerations generally yield quantitative criteria

to identify and dispense with discretization schemes that are only marginally
more accurate but require substantially larger computational resources for a
solution.
Closely associated with the above notion of computational economy are
considerations related to variable storage requirements. Although significant
advances have recently been made in extending the storage capability of current machines and the trend is likely to continue in the future, providing
adequate storage to solve practical engineering problems, especially in three
dimensions, still remains a major problem even for non-reacting flows. Considering the enormous additional complexity introduced by reaction, to obtain
sufficiently resolved solutions for the two-phase, turbulent, reacting, radiating flow in arbitrary combustor configurations places prohibitive demands on
the storage available. Thus, in spite of the extended finite difference computational molecule (hence storage) implied by more accurate discretization
schemes, their use might still be preferred over schemes of lower accuracy as
the grids required may be relatively coarse.
Finally, some additional requirements of a more practical nature are
adopted by the present study to arrive at rational choices for the schemes
considered. These pertain to implementation and application details and
specifically refer to compatibility with the present codes and the inherent
complexity. As the present study was structured to develop improved discretization schemes and solution algorithms for use with TEACH-type methodologies,
it is important that the use of proposed schemes be compatible with the constraints implied therein. In addition, generally speaking, the more complex
the scheme, the more difficult is the implementation and application and hence
possibility of errors, especially in imposing boundary conditions in a finite
duration program.
The following list, based on the above discussion, summarizes the
desired attributes that potentially viable techniques should possess to merit
further consideration:,
i) Accuracy
ii)
Stability
iii)
Expected efficiency

-38-

iv)
v)
vi)

Variable storage requirements

.
>
.
,, :. .Expected computing times
,.
:.
Ease with which the techniques can.be incorporated into existing
3-D codes for computing turbulent, recirculating flows.
It is clear from the above list that strict satisfaction of all the
items considered is exceedingly difficult, if not impossible. Thus, invariably judicious compromises have to be introduced. It can be argued that the
most significant requirements regarding discretization schemes considered are
that, they are conservative and bounded, as these characteristics generally
enhance accuracy. Also a bounded scheme yields a well-conditioned coefficient
matrix which is diagonally dominant, irrespective of the magnitude of the
Peclet number and the angle of inclination between the flow and the mesh.
Such desirable attributes are strictly realized for certain simple five-point
schemes like Hybrid or Upwind differencing schemes utilizing a finite volume
formulation. However, for the majority of the schemes with extended computational stencils considered in this study, issues of conservation and boundedness cannot be established a priori and qualitative/quantitative guidelines
have to be established to examine their expected behavior. In the present
effort, for this purpose a judicious combination of quantitative/qualitative
assessment practices have been employed. These include accuracy and linear
stability of the resulting difference equations via evaluation of the
properties of the coefficient matrix, Taylor series analyses and existing
heuristic stability analyses for iterative solvers commonly used in segregated
solution procedures like TEACH. These analyses, performed primarily for model
problems with known analytical or "exact" numerical solutions, yield further
qualitative information regarding solution cost effectiveness, complexity and
variable storage requirements. In addition assessment can be made of the
inherent accuracy and stability characteristics of the schemes considered.
Naturally, if a consistent finite volume discretization framework is employed
in the above test problems, conservation is strictly ensured locally, as well
as globally and the specification of boundary conditions arises naturally.
The above model problems incorporate highly idealized flows in the
presence of a dominant influence (convection, diffusion, source). However,
for practical performance evaluations such problems have to be replaced by
more realistic flow cases to examine the interdependency and delicate balance

-39-

displayed by coupled, nonlinear flow equations. Invariably, identification

and effective treatment of accuracy issues in practical engineering computations is a much more difficult but deserving undertaking and should utilize
effectively the results obtained
from individual model problems.
)
4.5

CLOSURE

This Section has reviewed in some detail the relevant characteristics of

accurate discretization schemes as well as the computational details of improved solution algorithms for incompressible flows. Practical criteria to
guide the identification and assessment of select techniques were established. These will be used readily in the next Section to provide a unified
qualitative/quantitative basis to discuss the pertinent details of the schemes
considered and hence examine their suitability for further quantitative
evaluation.

-40-

5.0 DESCRIPTION AND PRELIMINARY ASSESSMENT OF CANDIDATE SCHEMES x 5.1

IMPROVED DISCRETIZATION AND RELATED TECHNIQUES

The following potential schemes were Identified and qualitatively

assessed to provide improved (more accurate) predictions of isothermal, incompressible, turbulent, recirculating flowfields:
1)
Second Order Upwind Differencing Scheme
ii)
Advanced Skewed Upstream Differencing Schemes
iii)
Higher Order Schemes including Various Compact Implicit Differencing
Schemes
'
.
iv)
Explicit Dissipation Schemes
v)
Improved Flux Blending Schemes
vi)
Variational Discretization Schemes including Finite Elements
vii)
Spectral Methods
viii)
Lagrangian Methods
ix)
Adaptive Gridding and Modified Equation Analysis
The following sections briefly describe these techniques with references
to original works and comment on their relevancy, strengths and shortcomings
relative to the application of interest. The assessment criteria developed in
the previous Section will be referred to frequently to identify potential
techniques for subsequent quantitative evaluation in a two-dimensional code.
Although the structure, application and other attributes of the considered schemes vary widely, a careful scrutiny of the relevant literature
reveals considerable success in meeting the primary concern of accuracy for
most schemes in problems of interest. Hence, this preliminary assessment .wi 1.1,
emphasize Issues of accuracy, and stability leaving the less significant criteria of economy, compatability, tetc. as practical constraints in the
selection of the final four schemes.
5.1.1

Second Order Upwind Differencing Scheme (SOU)

This scheme is going to be discussed 1n detail in Section (6.2), however, some of its relevant characteristics will be outlined below for completeness.

-41-

Second order upwind Differencing Scheme (SOU) combines, in general, the

desirable monotonic character of upwind schemes with second order accuracy for
certain grid distributions. The scheme was first introduced by
Atias, et al ' in model problems including the square cavity problem and
the impinging jet flow. The convergence and stability characteristics of the
(29)
scheme were subsequently studied by Gupta and Hanohar
for the linearized
one-dimensional vorticity transport problem. During Phase I Aerothermal
Modelling effort sponsored by the Nasa Lewis Research Center, it was further
(2)
studied by General Electric
in a variety of model problems.
The stability analysis, of the scheme, via a one-dimensional scalar
transport problem, reveals that the characteristic difference equation exhibits a monotonic behavior with no spatial instability for uniform sign of convecting velocity. However, the case of the velocity changing sign across a
finite volume cell, characteristics of recirculation boundaries, warrants
further investigation.
The scheme can be shown to be second order accurate only on a smoothly
varying grid using an error analysis provided a Taylor Series Expansion,
Section (6.2). However, owing to the introduction of an extraneous boundary
condition in the solution of the model difference equation (three solutions vs
two required by the differential form) care needs to be exercised in the proximity of the boundaries for uniform application throughout the domain. In
(29)
fact, it is claimed by Gupta and Manohar that the overall accuracy of
the scheme is severely affected by the inconsistencies introduced in boundary
condition specification.
5.1.2

Advanced Skewed Upstream Differencing Schemes

These are going to be discussed in detail in Section (6.3), however a

brief description is provided here for introduction.
(12)
The original Skewed Upstream Differencing (SUDS) of Raithby/ '
although non-diffusive, was for some applications, plagued by problems of
unbounded solutions. Furthermore, potential iterative solution difficulties
were evident, as the resulting coefficient matrix was not diagonally dominant
owing to the possibility of main model coefficients becoming negative, e.g.
a in equation (3.23). A practical means of overcoming the boundedness and
/5\
related issues was later introduced by Syed, et alv ' who, blending "SUDS
with a less accurate but bounded scheme (e.g. Hybrid differencing) in varying

-42-

ratios, was able to obtain improved solutions. However, the particular form
of the adopted blending strategy is not strictly path independent, thus ,,:
possibility of non-unique solutions exist.
Recent works by Huget
and Raw
using similar implementations
of the Physical Advection Correction (PAC) concept within the framework of
SUDS, have overcome most of the original objections to the scheme as well as
providing a sound formulation to replace/augment the blending strategy. Two
variants of advanced SUDS are considered:
i) Linear Profile Skewed Upstream Differencing Scheme (LP-SUDS): admits
minor oscillations in the solution, but consequently yields most accurate results (uniformerly second order) of any SUDS scheme for a variety
of problems.
ii) Mass Weighted Skewed Upstream Differencing Scheme (MW-SUDS): a positive
definite skew scheme that produces stable, bounded solutions which are
significantly more accurate than those Hybrid differencing.
As with SOU, these schemes can be formulated conveniently, in a manner
to satisfy most of the assessment criteria.
5.1.3

Higher Order Schemes Including Various Compact Implicit Differencing

Schemes

These will be.discussed in detail in Section (6.4), however the following serves as a brief introduction.
Use of higher order schemes (fourth and up) in atmospheric and oceanographic applications has been advocated primarily by Kreiss. (31) Order in
this context strictly refers to classical error analysis provided by a Taylor
Series Expansion, Section (3.3.3)..;
The advantages of fourth order schemes over second order schemes are
most aptly demonstrated for the test problem of the one-dimensional wave
(hyperbolic) equation. Using centered second order differencing and a widely
available fourth order approximation to the advection term, Orszag (32)
examined the structure of local errors in the difference scheme and concluded
that the fourth order scheme requires roughly a factor 2 less resolution to
achieve a 5 percent maximum error, than does the second order scheme.
Furthermore, the fourth order scheme also localizes the error near the singularity better. The second and fourth order schemes used in this and other
studies are due to Arakawa. (33) They possess extremely desirable properties

-43-

in identically conserving vorticity, enstrophy (half the vorticity squared)

and kinetic energy. Due to these characteristics, they are especially well
suited to study hydrodynamic stability problems. Since they conserve
(34)
enstrophy they are not subject to the nonlinear instability of Phillips
/ I Q\
which arises from aliasing errors.
Among the objections raised towards the use of higher order differencing
schemes in complex flow field simulations, one notes, in general, issues related to cell Reynolds number restrictions (boundedness and transport!ve
characteristics), cost and the nature of the solution algorithms. For
example, the conventional fourth order differencing for the first order advection term involves five points in the formulation. This practice generally
suffers from the usual problems associated with the degradation of numerical
accuracy at the boundaries (an excessive number of image points have to be
used for a uniformly valid application of the technique). Furthermore, the
solution of the algebraic differencing equations is accomplished by more
involved steps than the application of the simple Tri-Oiagonal Matrix
Algorithm, Section (6.4.4).
There is a certain class of higher order schemes which have been shown
to resolve some of the deficiencies associated with the conventional schemes.
These are generally referred to as Operator Compact Implicit (OCI) schemes.
Related techniques including Hermitian differencing, Pade1 approximations,
spline collection techniques of Rubin and Khosla (35) can essentially be
developed in the same spirit as OCI schemes. However, experience with these
is relatively limited. Thus the following discussion will concentrate on OCI
schemes.
Classical OCI schemes are designed to produce a relationship between the
unknown solution t and the entire differential operator L(0) on three adjacent
points using a general finite difference approach. If the resulting scheme is
formally fourth order accurate, then the method is known as OCI. Note that
the formal fourth order accuracy is the highest that can be obtained by such
derived methods, unlike the spline collection schemes of Rubin and
KhosU.'35'
Experience with such schemes using model and practical problems/has
revealed serious issues of stability and boundedness as well as degradation in
accuracy with grid non-uniformity and implementation of non-physical boundary
/ Ot \
conditions, Ciment.
To overcome such underslrable attributes
(37)
Berger, et al
recently derived a further class of OCI schemes. These
consider the OCI coefficients obtained from the Taylor series approach as

-44-

asymptotic series in the mesh size. The resulting scheme contains nine free
parameters. Using six of these parameters it is possible to obtain formally a
fourth order method with no cell Reynolds number limitation. Moreover, th.e ,
tri-diagonal system of equations is diagonally dominant. The fourth order
accuracy is normally achieved by incorporating an appropriate exponential
character into the coefficients. However, the cost issues involved in the
iterative computation of such exponentials need to be examined critically^
Section (6.4.3).
.....
,
5.1.4

Explicit Dissipation Schemes

The idea of artificial dissipation is credited to Von Neumann and

/ ^gx
Richtmeyer
in connection with shock capturing calculations. This concept is very commonly used with time marching, compressible, Lax-Wendroff
and/or other implicit schemes to attenuate high frequency oscillations associated with shocked flows. The essence of the method lies in introducing an
artificial dissipation term into the difference equations (not the differential equation) such that the added term is of order higher than the truncation
error of the difference scheme provided by a Taylor Series Expansion. The
dissipation so introduced, however, does not represent any loss of energy or
other conserved quantities; the kinetic energy that is removed from high frequency oscillations is left in the system as internal energy of the fluid.
Most explicit dissipation schemes utilize pressure as the relevant sensor to activate the dissipation term (appropriate for shocked flows),
MacCbrmack. (39) For uniform pressure (or slowly varying pressure, characteristics of the flows considered here) other variables have to be considered
to serve the same purpose. Advantages of explicit dissipation schemes are the
strict local control of the boundedness and stability of solutions provided by
an inherently dispersive but accurate scheme (Central differencing), as well
as the simplicity of the corresponding formulation. Moreover, such schemes
have seen an extended development and application period in the prediction of
mostly compressible external flows using time marching techniques by
Jameson,
Beam and Warming,
MacCormack,
etc. However, the
somewhat non-unique nature of the implementation for various problems, as well
as the influence of the added terms on the solution, in conjunction with the
implicit dissipation provided higher order upwind schemes are only now
beginning to be analyzed for time dependent implicit schemes, Pulliam. (43)

-45-

(41)
The time dependent Beam and Warming algorithm
which requires a ':
fourth order damping term when spatial Central differencing is used, was
(44)
recently adopted by Kwak, et al
to obtain numerical solutions for incompressible Navier Stokes equations in tf generalized system. These authors used
(45)
the artificial compressibility technique of Chorin
to perform a time
integration of the equations in primitive form. The formulation of the artificial dissipation term was based on a fourth derivative of the form
(x-direction):

4
S. = - a/8 (AX ) (^-f)
= - a/8 (0.,., - 40. . ' + 60. - 40. , + 0. .)
J
J
J
J
3 d
J
3X* j
.
*
'
"'
~
(5.1)

where a is an adjusted constant, 0 is a conserved variable, j is the particular grid node. A similar expression is used in the y-direction.
An important point regarding time dependent methods must be further
elaborated. Convergence to steady state can be characterized (based on the
theory of characteristics) as a physical process of acoustic and entropy waves
clearing from the computational domain. A spatial smoothing procedure can
then be viewed to cause dual effects; dissipation of transient acoustic/
entropy waves and dissipation of high frequency oscillations introduced by the
differencing schemes (oscillations are real solutions of difference
equations). (18) Artificial dissipation introduced into difference equations
as a term of higher order than the truncation error of the scheme is unlikely
to influence the true solution of the differential equation (either transient
or steady state). This means that enhancement of damping transient acoustic/
entropy waves would require a dissipation term which is either of the same or
lower order than the truncation error. The converged solution, however, could
be in serious error in such cases. Based on this argument it is readily seen
that the dissipation term must be designed to achieve stability of the, difference scheme (e.g. Central) by filtering growth of the high frequency oscillations.
It is relevant at this point to discuss the general nature of a dissipation term that could be used in relaxation algorithms (steady state discretization of equations) for modelling of incompressible, turbulent, recirculating flows.

-46-

A steady state matrix equation of the form (cf. equation (3.33)')':

A(0) = b

(5.2)

is replaced by an equation of the form:

A(0) - D(0) = b

(5.3)

where A denotes the, spatial discretization operator for the variable 0,

defined by a Central or an equivalent scheme, b the source and D the dissipative operator. D is assumed to be fourth order accurate with a form given for
two dimensions as:

D(0)

= D (0) + 0(0)
y

where

(5.4)

D (J)) =

D ( )=d

y'

i,;M/2-diJ-i/2

and

In the above i and j denote grid indices and K.

. is generally a
constant for model problems derived from a stability analysis of the linear
difference equations. For complex flows however, no rigorous stability
analysis exists and the spatial variation of K due to non-homogeneous,
non-linear flow processes has to be specified iteratively. A.
is a
suitable difference operator that renders the smoothing term fourth order
accurate, i.e.,
<5'5)

-47-

The formulation outlined above is by no means unique, however it introduces dissipation for the conserved variable in a consistent manner and has
(40)
(44)
been implemented successfully by Jameson,
Kwak
and Beam and
(41)
Warming
for damping spurious oscillations in rapidly varying regions.
Care has to be exercised in treating the continuity equation to maintain
proper accuracy using such techniques.
5.1.5

Improved Flux Blending Schemes

The objections associated with the semi-heuristic formulation and application of explicit smoothing terms in the difference equations can, in principle, be alleviated by schemes that add and subtract dissipation based on the
local behavior of the solution. A scheme with some of these characteristics
was first proposed by Boris and Book
and later generalized by
(47)
Zalesak
for a more general class of problems. The basic idea incorporated in such schemes is to construct the net flux at a point as a weighted
average of the flux computed by a low and high order scheme. The weighting is
introduced in a manner that ensures maximum use of the higher order scheme
while producing tolerable or no oscillations. The resulting procedure is commonly referred to as a flux correction scheme.
The nature and elimination of dispersive post-shock oscillations in a
shock capturing scheme have been the subject of intensive research effort
(38)
since the early attempts by Von Neumann and Richtmeyer.
There have been
numerous formulations with various levels of success to implicitly or explicitly filter out the oscillations encountered, e.g., upwinding, artificial
damping, flux correction and more recently flux vector splitting.
Under
NASA sponsorship some of these concepts have been evaluated for flow problems
similar to the ones considered here, by Syed, et al.
These authors
effectively used a weighted mean flux blending scheme incorporating first
order Upwind and SUDS schemes, to obtain bounded solutions of comparable
accuracy to those yielded by SUDS alone. The nature of the shortcomings
associated with the use of such blending procedures is discussed in
Section (5.3).
Notwithstanding some of the unresolved problems and implementational
difficulties, a similar flux blending strategy could be developed to exploit
the high accuracy provided by an appropriate fourth order differencing scheme,
e.g. OCI. Using such an approach it is also conceivable that some of the

-48-

obvious shortcomings of the particular blending formulation (source influence,

path dependency, etc) could be alleviated. In addition, flux blending provides a further basis for rationally constraining the artificial dissipation
coefficient of the previous section to obtain bounded, stable solutions to
equation (5.2).
5.1.6

Variational Discretization Schemes Including Finite Elements .,

The cell Reynolds number problem of the classical finite difference

methods leading to boundedness and stability problems could, in principle, be
relaxed by Galerkin based methods, or more generally speaking, by methods
founded on a variational principle that minimizes some error norms. Techniques employing Galerkin or other related procedures preserve certain
important integral constraints, e.g. total kinetic energy. A conserved quadratic integral of energy is necessary for stable integration of the equations
of motion by prohibiting unbounded nonlinear instabilities which often arise
with finite differences.
In the literature, variational techniques, including finite elements,
are presented as an alternative approach: for reducing the relevant differential equations to an algebraic set (finite volume, finite differencing, etc),
and hence do not strictly constitute distinct discretization methods. However, they are reviewed here briefly for completeness. Furthermore, recently
their use in becoming more frequent in similar problems of interest in this ,
study and they encompass a semi-mature field rich in concepts and applications
and hence can be exploited, Raw.
Variational methods, in particular finite element method, seek an
approximate solution to the differential equation over finite subregions or
"elements" in, terms of a piecewise-continuous local interpolation formula
(trial or shape functions), by controlling the error in some average sense
through a variational principle or Galerkin procedures of weighting. The
shape function is defined over an element, with elements joined together to
cover the entire domain. For flow problems Galerkin procedures are generally
used.
The finite elements may be of orbitrary shape, although commonly triangular or quadrilateral elements are based. The solution 0 is approximated
over each element by:

-49-

"' '"*~*

(5.6)

where denotes summation over all the grid points within the element of
interest and the N are the shape functions, defined piecewise element by
element. They are usually polynomials of low degree. The polynomial
coefficients are determined by the requirement of continuity of 0, but not
necessarily of its derivatives, between the elements. Thus if the differential equation describing the problem is represented by:

L(0) = 0

(5.7)

then using the weighted residual method in which the weighting function is
equal to the shape function, the Galerkin representation for the problem is
given by:
/ NJ L ( N

d R=

(5>8)

where R is the domain. By choosing a succession of weighting functions, a

system of algebraic discretized equations can be obtained.
It can be shown that the conventional Galerkin finite element method for
convection-diffusion problems experiences similar difficulties as the disper(49)
sive finite volume/difference techniques, Gresho and Lee.
Thus/means
have to be introduced to bound the solution using practices similar to the
techniques considered here. To overcome such problems, several upwind type
finite element methods including the Streamline Upwind Scheme
(similar
in concept to SUDS) have been proposed. The difficulties associated with such
formulations, either in terms of unbounded solutions or excessive diffusion,
are similar in nature to finite volume/difference methods. However, recently
Raw
developed a co-located control volume based finite element procedure. The approach taken by the author introduces a rational basis for
approximating the convected variable at the control volume face and provides
the basic concepts involved 1n the development of advanced SUDS,
Section (5.1.2). Raw relates the interface variables to the nodal variables

-50-

via an algebraic approximation of their respective differential equation

applied at the control volume surface. This causes, .for flow problems* the
appearance of pressure in the continuity equation in a form that removes the
need for "reduced order pressure interpolation", typical of most finite element methods. In addition, a discrete convection operator is developed that
does not produce non-physical spatial oscillations and does not suffer ;f%rom
false diffusion.
- - . There is a further very desirable attribute of finite elements for flow
problems in the flexibility and generality provided by various element ,
shapes. .Such elements might also be moved appropriately to track interfaces
and fronts for improved resolution.
5.l.'7 Spectral Methods
Spectral methods are based on representing the solution to a problem as
a truncated series of smooth functions of the independent variables. Whereas
finite element methods, discussed above, are based on expansions in local
basis functions, spectral methods base expansions in global functions. Spectral methods are the extension of the standard technique of separation.of
variables to the solution of arbitrarily complicated problems.
Spectral methods involve the approximation of the flow using appropriate
combinations of Fourier series, Chebyshev polynomial series, Legendre polynominal series, etc.
While these methods do guarantee infinite order rates
of convergence to smooth flows, they appear at first to be excessively complicated compared to more conventional low order difference methods. However,
with the introduction of transform methods and spectral iterative methods,
they can be applied and used effectively to adequately resolve complicated
flow physics in arbitrary geometries. Spectral methods offer the significant
advantage over other (lower-order) methods in that they automatically achieve
high order accuracy at rigid no slip boundaries.
The underlying concept behind the transform methods, introduced to deal
with the nonlinearities present in the equations, Gottlieb and Orszag
(52)
and Gottlieb, Hussaini and Orszag
uses spectral representations only to
obtain accurate evaluations of derivatives and to impose the boundary conditions accurately. All the complicated nonlinear physics is evaluated locally
in physical space, so that flow physics of any desired degree of complexity
may be accommodated.

-51-

The extension of spectral methods to complicated geometries was investigated using two key ideas by Orszag
and Gottlieb, et al:
the spectral iteration technique and spectral element patching. These will be
described briefly.
Spectral approximations to general boundary value problems lead to full
NxN matrix equations for the N expansion coefficient an . It would seem that
3
the solution of these equations requires 0 (N ) arithmetic operations, while
2
storage of the matrix requires 0 (N ) memory locations. Since typical problems involve N ~ 10 , the direct solution (or even the direct formulation)
of such problems is clearly unworkable presently. The spectral iteration
technique circumvents this problem by allowing the solution of the full matrix
problems using just the storage ana" computational work required by low order
finite difference methods. The underlying idea is to approximate the spectral
operators that must be inverted by suitable sparse matrix operators and devise
an iteration scheme (in a deferred corrector sense) that leads to machine
accurate solutions in only a few iterations.
Consider the solution of a general linear differential equation L0 = b.
Let an N term spectral approximation to this problem be given by:
L

s P *N = b N

(5 9)

'

where b.. is a suitable N term approximation to b. Equation (5.9) is generally a full NxN matrix.
Suppose it is possible to construct an approximation Ldp to i the spectral operator L that has the following properties.
i) Lap has a sparse matrix representation so that it can be represented using only 0(N) storage locations.
ii) Lap is efficiently invertible in the sense that the equation

VN ' "N

<5-10'

is soluble as efficiently as a first order finite difference approximation to

the problem.
iii)

Lap approximates Lsp in the sense that,

L

ap

sp

-52-

for suitable constants m, M as N - . Roughly speaking equation (5.11) requires that the eigenvalues of Lap Lspn be bounded from above
, and below
as N ->. . Examples regarding the necessary bounds and other convergence con(53)
siderations for some model problems 1s provided by Orszag.
In this manner the solution of (5.9) can be accomplished using order N
storage locations with the number of arithmetic operations of the order of
N log N and the number of operations required to solve the above equation by a
first order finite difference method. The important conclusion is that spectral methods for general problems, in arbitrary geometries, can be implemented
efficiently with operational costs and storage not much larger than that of
the simplest finite difference approximation (e.g. first order Upwind) to the
problem with ,the same number of degrees of freedom. Since spectral methods
require many fewer degrees of freedom to achieve a given accuracy than required by finite difference approximations, important computational efficiencies can result. Nearly equivalently, spectral methods achieve much higher
accuracy for a given number of degrees of freedom.
The second key concept involved in application of spectral methods in
complex geometries refers to patching together complicated domains out of simpler ones by using moderate sized "global spectral elements". Because of the
nice boundary properties of spectral methods, little accuracy is lost by
patching together these spectral elements. However, the patching scheme does
give essentially complete geometric flexibility, especially where the global
elements are isoparametrically mapped curved elements. The resulting spectral
patching/spectral element methods combine the efficiency and flexibility of
finite difference/finite element methods with the accuracy of spectral methods.
Application of spectral techniques in problems involving non-smooth
solutions using certain eigenfunction expansions, may result 1n undesirable
Gibbs 1 oscillations 1n the vicinity of the discontinuity. However, such problems are conveniently treated by either patching the solution at discontinuties or pre and post filtering the solution.
5.1.8

Lagrangian Methods

Although pure Lagrangian methods are not strictly viewed as distinct

discretization techniques, they possess a very appealing feature in that the
numerical formulation does not include the convective derivative, the major
source of difficulty in Eulerian calculations. Thus, Lagrangian calculations

-53-

are able to treat high Reynolds number flows for long times relative to their
Eulerian counterparts and can effectively treat interfaces, free surfaces and
moving boundaries. However, the methods do suffer from severe drawbacks in
practical calculations. The foremost problem occurs in multi-dimensional calculations when the convection of the grid with the flow leads to large grid
deformations and a corresponding decrease in numerical accuracy. Regrldding
and remapping techniques introduce severe numerical diffusion into the problem
just as if a simpler Eulerian calculation were performed Initially. Schemes
incorporating Lagranglan and Eulerian features Into a composite calculation
(54)
algorithm, e.g. ALE
suffer degradation as any relocation of the grid
necessarily re-introduces numerical diffusion through the smoothing provided
by interpolation.
A second problem arises because a high order of accuracy, particularly
of spatial derivative terms, 1s difficult to achieve in Lagranglan calculations. When the computational grid moves, uniform spacing is not generally
possible. To construct high order derivatives in a time varying non-uniform
mesh is difficult and first order algorithms are hard to avoid. One problem
of maintaining monotonicity, is replaced by another; the introduction of first
order aspects into the calculation.
Lagranglan calculations in multl dimensions are very complicated and can
be much more expensive per grid point than Eulerian computations. Another
problem is that inexpensive direct solution algorithms of elliptic equations
cannot be applied. Finally, adaptive griddlng
is just as important in
Lagrangian simulations for front resolution as it is in Eulerian calculations,
and is more difficult to Implement.
5.1.9

Adaptive Gridding and Modified Equation Analysis

Techniques considered in this Section contribute only indirectly in

obtaining improved resolution and other desirable aspects of solutions.
In general calculation procedures a priori spatial information concerning significant gradients of dependent variables is usually not available,
thus the practical grids used reveal wasteful practices and the Inability to
satisfactorily resolve regions of interest.
Adaptive griddlng provides a means of overcoming the above problems via
grids generated dynamically as the solution develops. The basic concepts have
been successfully used in boundary layer computations for years, to capture the

-54-

spatial and temporal details in a sound and economical way. For flame propa(56)
gation in a solid, Dwyer and Sanders
used the scheme in a very 'effective
manner to generate adequately resolved, economical solutions. Recently a
sound mathematical basis was formulated for general application by
Brackbill, et al
incorporating grid smoothness and skewness; together
with resulting cell volumes in a functional to be minimized.
The adaptive gridding technique coupled with an accurate numerical technique offers an indirect means of resolving the cell Reynolds number (boundedness) problem of current prediction techniques. The solution continually
updates the grid either explicitly or implicitly clustering mesh points in
high gradient regions to satisfy the critical Reynolds number criterion. A
potentially viable way of achieving sufficient grid resolution is through the
use of nonlinear modified equation analysis. (58) Apart from establishing a
sound basis for generating adaptive grids, the modified equation approach also
yields insight into the local behavior of the numerical scheme by providing
the relative magnitude of the truncation, errors to the original terms of the
differential system.
'
The modified equation is derived by expanding in a Taylor series the
difference equations of a particular numerical scheme. While the modified
equation contains an infinite number of terms with increasing powers of spatial mesh intervals, it does represent the system of original differential
equations solved by the differencing scheme. It is not possible however, to
investigate the complete modified equation. For dissipative differencing,
most of the information regarding nonlinear truncation errors is contained in
the leading few terms. Therefore it is sufficient to investigate the truncated modified equation. The important point is that the modified equation
gives the exact nonlinear truncation errors for the complete system of differential equations being solved numerically.
Since the higher order terms in the modified equation do not appear in
the original differential equations, contributions from these term lead to
inaccuracies when compared to an analytic solution of the original differential equation. The removal or cancellation of these terms results in a higher
order integration scheme. In fact, an essentially third order scheme results
by removing only the dominant second order terms. The improved accuracy 1s
achieved with no change to the differencing scheme and with very little extra
computational work.

-55-

Devising an automatic mesh clustering scheme is accomplished via the

local truncation error levels provided by the modified equation analysis and ..
the choice of a suitable smoothing filter. Here special care has to be exercised in resolving the stiffness introduced into the governing equation
through the grid refinement procedure. However, the impressive grid reduction
achievable, especially in three dimensions (one or two orders of magnitude),
for adequate resolution, lends support for developing such schemes.
5.2

IMPROVED SOLUTION ALGORITHMS FOR ENHANCING CONVERGENCE AND EFFICIENCY OF

METHODS FOR INCOMPRESSIBLE FLOWS

Currently most incompressible flow solution algorithms adopt a particular form of a general segregated solution procedure discussed in detail in
Section (3.4.3), SIMPLE, SIMPLER, PISO, SIMPLEC, SIMPLEX, etc. Central to all
the above approaches is the generation of one or more Poisson-type pressure/
correction equations that constrain the flow to satisfy the vanishing of the
mass divergence. As remarked in Section (4.3), the algebraic solution of such
equations accounts for a significant portion of the total cost of solving the
flow problem, hence efficient and economical solvers have to be devised.
These are presented briefly in the following discussion, prior to their
detailed development in Section (6.5).
Considerations regarding the structure of the discretized pressure/
correction equation (3.45) (symmetric, with properties similar to those for a
diffusion process) permits the grouping of various solvers as base solvers and
acceleration techniques. These are:
i) Approximate Factorization Techniques as base solvers.
These include, but are not limited to, Incomplete Choleski Algorithm and
Stone's Strongly Implicit Procedure
ii) Conjugate Gradient Acceleration
iii) Block Correction Acceleration
iv) Multilevel Multigrid Acceleration Schemes
Finally, coupled numerical schemes are considered briefly for completeness. These are much different in philosophy, structure and implementational
details than a general segregated solution procedure.

-56-

5.2.1

Approximate Factorization Techniques

A commonly adopted technique for solving linearized algebraic equation

sets such as (3.45) in multi dimensions is repeated application of an Alternating Direction Line Gayss-Seidel solver. However, uncertain convergence
rates associated with the use of such solvers in certain problems with
strongly anisotropic coefficients, mixed boundary conditions, etc necessitate
development and application of more implicit solvers. A particular class of
such methods are conveniently based on approximate factorization techniques.
For instance, Incomplete Choleski (1C) is a suitable approximation to the con(59)
ventional Choleski Decomposition.
In 1C, the Choleski Decomposition
matrices are approximated by some simple sparse diagonal matrices whose
coefficients are determined from constraints imposed by the product of the
approximate decomposition matrices. In fact, the constraints are such that
the product of the approximate decomposition matrices must be identical to the
product of the Choleski Decomposition matrices for every non-zero entry of the
latter product. Naturally, the product of the approximate decomposition also
contains non-zero entries where none should appear.
Stone's Strongly Implicit Procedure (SIP) is an alternative approximate
factorization technique: However, 1n SIP an attempt is made to partially cancel the effect of the spurious non-zero entries of the product of the approximate decompositions. A parameter, a, is introduced to control the degree of
partial cancellation. In fact, when the value of a is set to zero, i.e. no
cancellation, it can be shown that SIP and 1C become algebraically
equivalent.
Although as base solvers both 1C and SIP have often been shown to be
attractive relative to other iterative solvers, their rate of convergence is
still unacceptably slow in many cases. To enhance their rate of convergence, a number of acceleration techniques are available.
5.2.2

Conjugate Gradient Acceleration

Due to the symmetry of the linear set of equations for pressure, the
generalized Conjugate Gradient Method (CG) of Concus and Golub ' could be
used effectively for their solution. Employing this acceleration technique in
conjunction with 1C (or equivalently, SIP with no partial cancellation), the
resulting method becomes a finite (direct) method, with each iteration determining an additional orthogonal basis vector of the solution such that a pre-

-57-

defined norm of the error is reduced. Unfortunately, due to the asymmetries

introduced, the Concus and Go Tub variant of CG acceleration is not appropriate
for accelerating SIP with partial cancellation.
5.2.3

Block Correction Acceleration

The Block Correction acceleration (BC) is applicable to both 1C and SIP

(64)
methods. Based on the procedure of Settari and Aziz
and as discussed in
some detail by Hutchinson and Raithby,
BC technique adjusts the solution
in blocks, normally along rows or columns of the computational domain, by an
additive constant such that the residuals in each block sum to zero. A particular mode of application of BC could apply it independently along both rows
and columns with one 1C or SIP iteration employed between each row and column
BC. Specifically such a sequence is given by:
i) the application of BC to ensure that the residuals in each column sum to zero,
ii) an application of one 1C or SIP iteration,
iii) the application of BC to ensure that the residuals in each row
sum to zero, and
iv) the application of one 1C or SIP iteration.
In steps i) and iii) given above, the Tri-Diagonal Matrix Algorithm can
be used to solve the BC equations.
5.2.4

Multilevel Multigrid Acceleration

In algorithms incorporating "explicit" schemes like the line Gauss

Seidel or Successive Line Over-relaxation for the solution of the algebraic
equations, the high frequency error modes are reduced very effectively. However, the attenuation of the low frequency components is very poor. The
Multilevel Multigrid technique attempts to resolve this dilemma by employing a
sequence of grids ranging from fine to very coarse to eliminate the different
error modes. As such, it can be viewed as a routine for reducing or smoothing
out the error modes in a particular frequency band on a given grid structure.
For relaxation type solutions, the error vector at any phase of the
iteration cycle is expressable as a discrete Fourier transform. If a small
number of iterations are subsequently carried out using a fine grid, the high
frequency modes of the error are much reduced, although the low frequency end
of the spectrum remains largely unaffected. If recourse is then made to a

-58-

coarser grid, say with half the number of grid points, due to the resolution
capabilities of the grid, the solution will only reduce error modes of correspondingly lower frequencies. This cascade-like process is continued down to
the coarsest grid until the complete error spectrum has been covered. There
exist various applied versions of the above ideas collectively referred to as
/ to\
classical multigrid methods, by Brandt.
The following discussion will
emphasize the details of an alternative multigrid strategy based on BC ideas
(62)
and successfully used by Hutchinson and Raithby,
the Additive Correction
Multigrid (ACM).
In the ACM, coarser grids (contiguous blocks) are formed from 2x2 configurations of fine grid nodes, (two-dimensional), hence the solution of the
coarse mesh equations and subsequent correction for the fine mesh solution
yields a residual field that sums to zero over each block. In addition to
prescribing the method used to evaluate coefficients on the coarser meshes,
the ACM method also prescribes the manner in which residuals are transferred
to the coarser grids, how coarse grid corrections are applied to the fine grid
solution as well as how boundary conditions are treated. It is also worthy to
note that ACM is equally applicable to symmetric and non-symmetric systems of
linear equations and is not restricted to fine grids with 2 x2 points in
the two coordinate directions.
A particular application of the above methodology reproduces the classi(63)
cal "flexible cycle" multigrid algorithm of Brandt
with one difference.
Instead of iterating on the fine grid until convergence becomes slow, only one
fine grid iteration is performed before initiating the solution on the coarser
mesh. Employing IC/SIP on each of the meshes (ranging from the finest to the
coarsest, a 2x2 mesh in two dimensions),ithe single Iteration residual reduction rate for each of the meshes except the finest and coarsest may
conveniently be chosen to be 0.5. Regarding the specific details of terminating the smoothing procedure, iteration on all but the finest and coarsest mesh
is considered sufficiently converged when the 1 -norm of the residuals is
less than 25 percent of the 1 -norm of the residuals when they are last
obtained from the next finer mesh. Finally, iteration on the 2x2 mesh is
taken to be converged when the 1 -norm of the residuals 1s reduced to
0.01 percent.

-59-

5.2.5

Coupled Numerical Schemes

Conceptually, the elliptic influence and interequation coupling in a

computational domain can be properly accounted for by discretizing all the
governing equations at every grid point and then simultaneously solving them
using coupled iterative solvers.
For three-dimensional applications
especially, such solution procedures require computational resources that are
beyond the capabilities of current machines. Another approach deserving study
is to use parabolic solutions with "best possible" elliptic corrections (with
appropriate treatment of the continuity equation). SIMPLE/derivative
algorithms belong to this class of solutions. These methods, however, require
the use of variable under-relaxation parameters to obtain a converged solution. As methods progress in complexity to account for elliptic influence,
the need for a lower under-relaxation parameter decreases. SIMPLER thus
(64)
requires a lower under-relaxation parameter than SIMPLE.
Another
example of such an approach is reported by Caretto et al,
where the
authors adopted the SIVA (Simultaneous Variable Adjustment) procedure. This
is basically a point iteration method which properly couples the continuity
and momentum equations (to satisfy mass conservation better) and obtains all
dependent variables simultaneously. The resultant scheme, although more
expensive due to the adopted point iteration procedure, did not require any
under-relaxation, while for the same problem SIMPLE was unstable requiring
under-relaxation.
An analogous observation was reported by Srivastava et al
in their
study of viscous shock layer equations. Viscous shock layer equations are
parabolic in nature with ellipticity introduced through the unknown shock
(67)
shape.
In their study it was found that any attempt to solve the
equations using parabolic procedures (with corrections on the shock shape
analogous to pressure corrections) required problem dependent underrelaxation. To alleviate the need for practices using ad-hoc specification of
under-relaxation parameter, a solution scheme was devised that couples the two
relevant first order equations and solves them simultaneously using either a
box or a staggered scheme.
All the above observations for the model problem are relevant in consideration of the structure of SIMPLE/derivatlve algorithms. Therefore, a
procedure that treats fully elliptic difference equations by approximate parabolic methods merits further consideration in the present study.

-60-

The concept of coupling, and simultaneous solution for the two first
order partial differential equations of continuity and radial momentum, of
relevance in the study of viscous shock layer structure, is demonstrated next
by considering the inviscid perturbation equations of compressible flow, i.e.,

where A = (1-M2) and M denotes the Mach number. Solution method depends on
the sign of A; the equations are hyperbolic for M>1 , while for M<1 or M=0, the
equations are elliptic.

Considering elliptic equations, the above equations can be coupled using

/ cg\
Keller's box scheme
extended by applying the difference equation in two
sweeps; x and y directions. Using Central differencing scheme, equations
(5.12) and(5. 13) applied at (j-l/2,k+l/2) and (j-l/2,k-l/2) , where j.and k
denote grid locations in the x and y directions respectively, will contain
variables at (j,k), (j,k+l), (j,k-l), (J-l.k) and (j-l,k+l). However, these
equations can be readily manipulated to obtain tri-diagonal forms for the x
and y sweeps in the following manner:
Applying the box scheme at (j-1/2, k-1/2) for equations (5.12) and at
(5.13)

*vj-i.k-vj.k-rvi-i.k-i>

=0

(v

j,k
(5.15)

where AX and Ay denote grid dimensions in x and y, equations (5.14) and (5.15)
can be combined to yield:

where D contains variables at (k-1).

-61-

Similarly equations written at (j+l/2,k-l/2) yield:

where D again contains variables at (k-1). Equations (5.16) and (5.17) can
now be manipulated to obtain a tri-diagonal form to readily yield x-direction
solutions for u. . and v. for a given k-line. y-direction sweeps to
determine U. . arid v
for a given j-line are accomplished in a similar
j,k
jt k
manner.
the above inviscid approach, although formulated using the box scheme,
can be extended to incorporate any discretization procedure by a suitable
elimination scheme. The key lies in realizing that an elliptic formulation
can be alternatively viewed as two parabolic formulations in two coordinate
directions, cf. SIMPLE/SIMPLER algorithms.
Treatment of the diffusion terms, present in general flow equations
could be handled in a manner suggested by Keller for a box scheme, or in any
heuristic manner for any discretization scheme as long as the coupled nature
of solutions for continuity and momentum are preserved. Furthermore, incorporation of viscous terms tends to stabilize the solution procedure and does not
contribute significantly to solution difficulties.
5.3

PRELIMINARY EVALUATION OF CANDIDATE TECHNIQUES

Due to the widely ranging nature of the techniques discussed above,

regarding issues of applicability, implementation and required computational
resources, a vigorous quantitative evaluation procedure is difficult to
device. In this regard the expertise associated with a Technical Advisory
Committee specifically established for this program, as well as a critical
assessment of information available in the literature was utilized fully to
arrive at rational choices. The Technical Advisory Committee included
individuals active in the field of computational fluid dynamics with the
members:
- Dr. S. A. Orszag
- Dr. S. G. Rubin
- Or. R. T. Davis
- Dr. C. J. Knight
- Dr. B. N. Srivastava

-62-

A close scrutiny of the candidate schemes considered, reveals a wide ^

disparity in relevant critical merits and shortcomings. Such issues pertain
to in/applicability (within the framework of TEACH methodology), extensive
further development and implementation, only marginal development over the
current algorithm capabilities, resources allocated to the present study.
Even though such issues were given only minor consideration in the foregoing
discussion that emphasized primarily accuracy, stability and convergence
enhancement, the impact they have on a finite effort has to be critically
assessed. The following discussion will thus provide mostly a qualitative
justification for the choices adopted in the study.
Use of explicitly damped schemes that incorporate a fourth order dissipation term in the transport equation and achieving higher order accuracy
via deferred correction techniques employing lower and higher order
(unbounded) difference operators, can be mathematically shown to proceed on
identical paths, i.e., second order accuracy can be recovered using first
order Upwind differencing for the convective derivative, if a correction term
evaluated at a previous iteration is included in the discretized equations to
give at convergence a Centred approximation. However, the artificial viscosity introduced in explicitly damped schemes has to be fine tuned to be
selectively effective in regions where it is required. Such a practice
strictly introduces issues of non-uniqueness and/or need for a priori knowledge of the solution. Furthermore, a rational basis for appreciating the
physical nature of the incorporated terms, in a manner similar to the Physical
Advection Correction (PAC) of the Skewed Upwind Differencing Schemes, (SUDS)
is lacking. Even though for simple model problems the impact of damping terms
can be clearly identified, for the complex viscous, recirculating flows of
interest here, the determination of the "optimum" dissipation coefficient and
the related issues of convergence and stability is usually determined in an
ad-hoc manner.
Flux blending schemes, that attempt to optimally combine the bounded,
but albeit smeared, solutions of a lower order scheme with that of an
unbounded, but more accurate higher order scheme suffer from similar problems. There have been utilized in the past, various formulations operating on
different solution variables in a tailor-made manner. Recently, two formulations blending first order Upwinding with SUDS have been assessed by
Syed, et al
who concluded that the final solution might not be path

-63-

independent. In fact, it has been suggested that different initial values of

the adopted blending factor may produce different converged solutions, where
convergence is measured by the level of residual source error. Even though
various arbitrary practices can be utilized to reduce or eliminate the resulting solution uniqueness, a scheme that incorporates the correct physical
influence in the the determination of blending factors is preferable and
should alleviate the reported solution difficulties.
Introduction of variational discretization techniques in complex, reacting flow field predictions (in the same vein as the established finite element
analysis of structural mechanics) is fairly recent and has not enjoyed comparable development and analysis as the corresponding finite volume/difference
methods. Even though the two formulations (finite volume/difference vs finite
element) can, in principle, be shown to be identical, and finite element techniques afford greater flexibility in the treatment of complex geometries,
practical issues of implementation, familiarity, maturity, etc. rule out their
being considered as serious candidates. In addition, there exist comparable
problems of convective differencing and the reduced pressure interpolation to
circumvent the checker-board effect of incompressible flows in most formulations.
Spectral methods enjoy enormous popularity in transition/stability predictions and their use is becoming widespread in a variety of applications.
However, for practical combustor computations dominated by high Reynolds number effects and various coupled physical processes,, their routine use requires
extensive further development and computational resources. Lagrangian methods
were already found unsuitable for the problems of interest here, due to
unresolved issues of rezoning and consequent numerical diffusion, implementation in three dimensions, etc.
Finally, adaptive gridding (and the related modified equation analysis),
in spite of the intense current activity and interest therein, has not
achieved adequate maturity to consider its use in reacting three-dimensional
flow situations and thus remain a research topic. Furthermore, for coupled,
nonlinear equations, e.g., full Navier Stokes, modified equation analysis
becomes extremely unwieldy and cumbersome.
Concerning the use of coupled numerical scheme/solution algorithms to
substantially enhance convergence, a critical assessment as well as
discussions with S.V. Vanka of Argonne Labs revealed serious potential draw-

-64-

backs associated with their use in the light of the criteria of

Section (4.4). In essence, these approaches are very much different in
philosophy, structure and implementational details than a general segregated
solution procedure, and thus require substantial departures from the methodology incorporated in TEACH. Furthermore, potential problems are anticipated in
three-dimensional applications due to required computational resources. These
and other related considerations (time and resource commitments for their
development being outside the scope of the present study) helped exclude them
from further consideration.
5.4 SELECTION OF FOUR SCHEMES FOR FURTHER QUANTITATIVE EVALUATION

Based on the above qualitative assessment, as well as considerations

regarding relevant additional information available in the literature and
prior experience, the following four most promising techniques were identified
as deserving further quantitative evaluation:
i) Second order upwind differencing scheme
.
t
ii) Variants of skewed upstream differencing scheme
iii) Variants of compact implicit method
iv) Strongly implicit procedure accelerated by:
a) Conjugate gradient algorithm
b) Block correction technique
c) Additive correction multigrid algorithm
It will be seen that the particular schemes selected address issues of
both discretization accuracy and convergence enhancement in a well-balanced
manner. Detailed formulation emphasizing issues of accuracy, stability and
convergence together with implementational details, including computational
requirements for the above schemes are provided in the next section. The
results of the two-dimensional evaluation exercises is presented in Section 7.

-65-

6.0
6.1

DERIVATION OF THE SELECTED SCHEMES FOR TWO-DIMENSIONAL EVALUATION

FINITE VOLUME DISCRETIZATION AND SOLUTION OF THE CONSERVATION EQUATIONS

FOR A SCALAR, MOMENTUM AND MASS

A general steady transport equation governing the conservation of a

typical scalar, momentum and mass can be expressed for both laminar and (timeaveraged) turbulent flows as, equation (3.19):

r (F0r

a? * S0

f6'1)

where 6=0 for two-dimensional (planar) flow and 6=1 for ax1symmetr1c flow.
The variable 0 represents any one of the dependent variables (e.g., the velocity components u, v, w, species 1, turbulent kinetic energy and dissipltation). The exchange coefficient, r0, represents the sum of both laminar
and turbulent contributions and is interpreted as effective viscosity for u,
v, w, the effective dlffusivity for species, etc. S0 denotes a generalized
source term. Table (6.1) presents the particular forms of S adopted for
0
various variables.
Using the techniques and practices developed in Sections (3.3.1) to
(3.3.3), equation (6.1) is integrated over a finite control volume appropriate
for each variable 0, figures (3.1) and (3.2), to yield the following conservation expression:
n

/*e

's ^xw

-66-

dxdy

(6.2)

TABLE 6.1 TYPICAL SORUCE TERMS FOR RELEVANT VARIABLES

Source Term S

Variable 0

X
ax'

ax

32
ax

1
r ar

rw

pw 2

ap

r ~ ar

k**

6-

f(mixture
fraction)
m. ***

Either Sf|| = S f u J

fu

(mass frac- 0 r Sf u =
tion of un- where S fu>2 = CRmfupc/k
burned fuel)
h
0 (except when radiation is present)

* ye denotes effective viscosity

** G is given by:

6=

au1
ax.

au. au,
ax.' ax.

i=l j=l
***

It is supposed that the rate of combustion of fuel per limit volume Is

governed by the Arrhenius-type relation ^u=SfU -j where F 1s a
pre-exponential factor and E 1s the activation energy. |[A,Bj denotes
the minimum of absolute values of arguments.

-67-

Adoption of Cartesian geometry for ease of highlighting the significant

aspects of the formulation is implied in equation (6.2) and throughout most of
this Section. In equation (6.2) each J., denotes the total local transport
of 0 across the i boundary of the finite volume by convection and diffusion,
while the last term signifies transport due to source generation/dissipation.
By taking the east boundary as an example, J 1s expressed as:
J

efr = Cefr + Def,

(6.3)

where

"
/

(Pu0)e dy

ys

and

rnn (r a*
t ff)e dy

Def = -J

(6.4)

ys

To approximate equation (6.2) for each finite volume by an algebraic

equation requires specific assumptions about the variations of 0 1n the integrals of equation (6.4) between the selected grid points in space. Making use
of the mean value theorem, that implies either constant or linearly varying
fluxes along each cell boundary, allows the Integrals to be replaced by:
c

ef = c e*e
.
Def - - de(ai/ax)e
where subscript e denotes the values of the associated quantities at the cell
interface. The subsequent step of expressing 06 and 30/ax)6 1n terms of
nodal values 1s crucial in that it Influences boundedness, stability and other
characteristics of resulting solutions and will be taken up 1n most of this
Section.
Introducing suitable approximations for 0 and 30/dx) as well as
the last term in equation (6.2), and subsequent linearization of momentum
equations yield the final algebraic representation of mass and momentum conservation as:

-68-

e UP * mw UW *mn VP * ms VS =

(6>6)

'

p u P = 'NPV- 0 " ( p E - p P > +b "

NP
'

Vp

NP VNP *C"(PN -PP>+ ^

(6;8)

NP

<6'7)

;::

,,..... NP denotes Influences of neighboring nodes. The various variables in

equations (6.6) to (6.8) have their usual definitions given in Section 3. The
absence of the dependent variable pressure in the mass conservation equation
requires that special numerical techniques be used to solve equations (6.6) to
(6.8), and many of these techniques are devised such that an equation for
pres- sure results. To minimize the computational requirements associated
with the formation and solution of the pressure equation, a segregated
approach is often adopted, Section (3.4.3). In the segregated approach (e.g.
SIMPLE) an attempt to minimize computational requirements 1s made by
introducing approxi- (nations to the momentum equations so that each nodal
velocity is expressed in terms of a local pressure difference:

up ^p - du (pE - pp)

Vp = Vp - dv (pN - Pp)

(6.5)

(6.1,0)

Substituting for the velocities in the algebraic representation of mass

conservation, equation (6.6), a pressure equation of the following form
results:

app PP NP

With equation (6.11) providing an equation for each nodal pressure, the resulting set of linear algebraic equations for the conservation of mass,
momentum and relevant scalars can be conveniently solved.

-69-

6.1.1

Calculation and Discussion of the Interface Fluxes

Jo provide a clear understanding of the significant aspects of discretization schemes in a unified approach, the following discussion will initially be developed in the context of the one-dimensional analog of equation
(6.1) and hence (6.2). Furthermore, where appropriate, reference will be made
to the specific details of the particular schemes presented in Section 5 to
establish a unified framework for assessment.
Equations (6.1) and (6.2) in one dimension become:
d_

d_

d0

'

/ A

A
0

d0x
0dxe

d0
0 dx w

.
0

where AV is the volume of the control volume. In equation (6.13) the interface derivative of 0 is conveniently approximated by employing the usual centred finite difference approximation characteristic of diffusion processes.
The representation of interface values of 0, however, is not as straightforward and the approach used distinguishes one discretization scheme from
another.
The discussion of various discretization schemes adopted in this study
begins by considering the Upwind Differencing Scheme (UDS) for u>0 (a quantitive baseline), where the east face value of 0 is taken as the upwind nodal
value of 0, i.e.,

'; '

The advantage of this approach 1s the simplicity and robustness of the

resulting scheme. Also in the absence of any significant diffusive or source
effects, the accuracy of this approach for one-dimensional flows is quite
acceptable. However, in the presence of diffusion or source effects, UDS 1s
only first order accurate and hence significant errors can arise. Therefore
it is necessary to improve the accuracy of equation (6.14). Recognizing that
equation (6.14) is the first term of a Taylor Series representation of 0 ,
the accuracy of equation (6.14) can be Improved by including additional terms
of the series. For instance,
-70-

AX

i
-

del

where ix is the distance between nodes E and P. Now if an adequate evalui

.,
ation of (d0/dx) can be found, the evaluation of 0 would be second
e
.
e
order. An obvious choice is the centred finite difference representation
about e given by:
/d0 > gE " ^P
dxe ~
*xi

(6.16)

Combining equations (6.15) and (6.16) and substituting for 0 as well

as 0 into equation (6.13), it can be shown that the resulting discretlzaw
tion scheme is Central Difference Scheme (COS). With this perspective CDS can
be viewed as UDS with the upwind representation of 0 corrected by a centred
difference representation of the spatial derivative. The disadvantage' of'CD'S
is the non-physical negative dependence of 0 on the node downwind of 0 ,
which is referred to as negative influence and which can result in nonphysical oscillatory solutions.
Naturally, there are alternatives to equation (6.16). One is to evaluate the derivative using an upwind finite difference representation,

Combining equations (6.15) and (6.17) and substituting into equation

(6.13) the result is the Second Order Upwind (SOU) scheme. The advantage of
SOU is that, unlike COS, negative downwind influences do not arise. Equations
(6.16) and (6.17) do not provide the only means of evaluating the derivative
in equation (6.15). In fact, a complete family of schemes result from taking
the following linear combination of upwind, centred and downward finite difference representations for (d0/dx) .

-71-

where a+b+c=l. The family of schemes described by equation (6.18) includes

(69)
CDS and SOU as well as the QUICK Scheme of Leonard.
The advantage of
equation (6.18) is that by appropriate choices of a,b and c, it is possible to
control the distribution of influence. Unfortunately, there appears to be no
combination of values of a,b and c which ensures that the resulting numerical
solutions do not exhibit spurious overshoots and undershoots.
Therefore
an alternative approach of evaluating (d0/dx) is considered.
The evaluation of (dfl/dx) given by equation (6.18) is equivalent to
e
assuming the variation or profile of 0 between the nodal values of 0 surrounding the east face of the control, volume. As a result, these schemes are
referred to as profile schemes. An alternative approach is based on physical
arguments. As mentioned previously, for one-dimensional flows the upwind
evaluation of 0 given by equation (6.14) is appropriate in the absence of
diffusion or sources. Therefore, the correction applied in equation (6.15)
can be interpreted as the correction necessary to account for the influences
of diffusion and sources. This physical interpretation can be arrived at by
first considering the non-conservative form of equation (6.12),

>u

< 6 - 19a >

'

or rearranging,

ii tfc <r. > *

Substituting for d0/dx in equation (6.15)
3x
<r

>*

"-20>

Clearly in equation (6.20) 0 is given by 0 and any changes in 0

that arise between the P node point and the east face due to diffusion and
source term influences. Consequently, the resulting scheme is referred to as
UDS with correction for diffusion and source term influences.

-72-

To complete the description of the physical correction, the evaluation

of diffusion and source term influences must be provided. There are a number
of possible methods available. A convenient choice is to base the evaluation
of the corrections on the evaluations already made in the control volume,
equation (6.13). For instance, based on upwind control volumes, the correction terms of equation (6.20) are given by:
.

6X.

i
2pu

r
L

d
dx

' d ^\
V1
0 d

S ]

x> + 0

6X.

40 \

-i. C 1

2pu

dx 'e

ltr

&)

0 dx' w

(6.21)

The correction terms can also be based on downwind control volume values, as well as a linear combination of upwind and downwind control volumes.
The physically based correction of the upwind estimate described above
uses an explicit evaluation of the correction terms. That is, the corrections
are expressed only in terms of the nodal values of 0. A more implicit
approach is taken in Operator Compact Implicit Methods (OCI), Section (6.4),
where the corrections are expressed implicitly in terms of operators and nodal
values of 0 which are both determined as part of the solution. The result is
an exponential scheme which accounts for convective and diffusive effects and
an implicit correction to account for source term influences.
In summary, for one-dimensional flows, the SOU scheme, UDS with diffusion and source term influences, and the OCI Scheme can all be expressed as an
Upwind difference scheme with correction to account for the variation in 0.
The SOU scheme bases the correction on profile assumptions, while the OCI and
UOS with diffusion and source term influences employ corrections reflecting
the effects of diffusion and sources.
The implications of the above one-dimensional analysis can be easily
extended to cover two-dimensional formulations. Approximations of interface
values can again be accomplished by Upwind differencing. However, for multidimensional flows use of UOS is accompanied with severe false diffusion, i.e.,
solutions display excessive amounts of smearing.
To eliminate false diffusion, a correction to the upwind estimate, as in
equation (6.15), is required. The correction can be determined from profile
assumptions, as provided by equation (6.18), or a physical correction which,
for two-dimensional flows, must account for the effects of diffusion and

-73-

source terms as well as a component of convection in one of the two coordinate

directions. For instance, the correction at the east face of a control volume
is given by:

<6 22)

For multi-dimensional flows OCI also requires correction, through the

appropriate definition of operators, for the effects of the additional components of convection and diffusion.
The advantage of the profile or physical corrections as described above
for two dimensions is that they are relatively straightforward to implement.
However, such implementations may also be subject to relatively large overshoots and undershoots, especially when flows are at a large angle to the
grid. The primary reason these overshoots undershoots arise is because the
corrections for the additional component of convection, based on either profile assumptions or physical arguments, introduce non-physical negative influence coefficients.
An alternative is to use Skewed Upstream Differencing Schemes (12)
(SUDS). In SUDS, the two convective terms of equation (6.1) are replaced by a
single streamline derivative term:
IT (P">) + fc (pv0) = pV f?

(6.23)

where V = \ u +v and s is the streamline coordinate. Upwinding along the

streamline, the convective terms at the east face of a control volume are then
given by:

(6 24)

'

where 0 is the skewed upwind value of 0 and L is the distance between the
locations where 0e and 0u are defined, figure (6.1). The correction of
SUDS can then be expressed as:

-74-

L{

l) e

where (30/3s) represents the correction required to account for diffusion

and source terms and is given by:

ax" (rei fx^ + fe-(r a?*+ s- ]-

(6 26)

'

Note that the correction to SUDS given in equation (6.26) does not include a correction for any component of convection. All components of conveo.tion are included in the Skewed Upstream differencing.
In summary, for multi-dimensional flows, SOU, SUDS and OCI schemes all
adopt an upwind representation of convection. Conventional implementations of
OCI and SOU employ upwinding along coordinate directions, while SUDS upwind
along streamlines. Furthermore, each of the schemes require correction of
Upwind differencing. In SOU corrections are determined from profile assumptions. In both SUDS and OCI corrections are more physically based with corrections in SUDS including diffusion and source effects, and corrections in
OCI including effects for diffusion, sources and components of convection.
With this background it is appropriate to consider the details of the
schemes.
6.2 SECOND ORDER UPWIND DIFFERENCING SCHEMES

6.2.1

Derivation and Characteristics of Finite Difference Equations

As discussed above, SOU schemes attempt to improve the convective

accuracy of UDS by the inclusion of additional terms in the Taylor Series
Expansion for the interface value. The scheme will be developed initially in
one dimension to study its accuracy, stability and convergence characteristics. Assuming a profile for 0 in the form given by (uniform convective
velocity and grid), figure (6.1),
0

P'
(6.27)

-75-

1-2

1-1

. I+ 1

I+ 2

NN

NW

tI

N
1I*

Sy,
^*j

ww

w,
ii

r
i
Wf
I
I

NE

Ax- l
n
I
I

-J + 1

rE

r EE

/K

J
[
0&
J
</^
_*

L__ s .

syj-i

^8xM

\*

J +2

sw

OlAf

_-

?Xi

cc
ot

,( SS

J - 1

J -2

N1659

Figure 6.1

Control Volume Details for SUDS (Positive Velocity at East Face)

and SOU

-76-

and similarly for 0w , the net convective flux balance is expressed as: ,

S *fc(f

(6.28)
u<0

Substituting for the convective derivative in equation (6.19a) with no

sources and incorporating the usual approximation for the diffusive fluxes
yields a difference equation for 0 . The stability and monotonicity aspects
of the difference equation are conveniently analyzed as follows (u>o): The
discretized form of equation (6.19a) is:

*2

i = 2,. ..n

In the above i denotes the particular node and P signifies the ratio
u/r . In addition to (6.29), the analysis requires the usual boundary conditions,
= 0

(6.30)

and the first order upwind formula to be applied at 1=1 to yield 0 , i.e.,

Denoting P4x by Pe, the grid Peclet number, and 0.-0.

equation (6.29) becomes:
Pe (3a. - a^^

- 2 (aui

-77-

- a.) = 0

by a , then

1-2

(6.31)

The boundary condition on a. is given by (6.30) as:

a

= (1 + Pe)a

(6.32)

Rearranging equation (6.31) yields

2 4^ - (2 * 3 Pe) a. 4
a three term, linear recurrence relation with constant coefficients.
there exist constants A,B, X , X such that
a = Axlj" + B X-1

where X and X

v = 1

Thus

n+1

satisfy
2 X2 - (2 + 3Pe) X + Pe = 0

(6.33)

i .e.,

= 1/2 { 1 + 2
f Pe + Wl + Pe + 9PeV4 [

(6.34)

Notice that Pe>0 implies X >0. A similar analysis yields identical

results for Pe < 0.
Use can now be made of equation (6.32) to obtain

( 1 t Pe) (AX1 + BX2)

or,

A [x,\ - (UPe)x.]
\ = -B [xi '- (l+Pe)xi0]
Now equation (6.33) implies

.. = Pe (-y-)

-78-

J - 1.2

a n d from equation (6.34)

.. ' . . . .

-| (x.j-1) > 0

and -| (\2-l) < 0

Thus, neglecting the degenerate case of A=0 and 8=0, AB>0, and hence a is
y
of a uniform sign (bounded).
.
.
Regarding the stability of iterative techniques used to solve equation
(6.29) it can be shown that such solvers, are well conditioned and converge
rapidly with a maximum eigenvalue of the iteration matrix assuming a value of
2/3. Furthermore, a formal Taylor Series Expansion yields the accuracy of the
scheme as second order on smoothly varying or uniform grids.
There have been proposed alternative formulations of equation (6.29) due
to the extended nature of the molecule implied therein.
. One such form
is:
3 u
1 u
(
I Ix" *i -*i-l> -2

(i

1-l - *i-2>

U>

(6.35)

f Ix"(-Wl> * 2 f c( i

* *>

u<0

The advantage provided by such a rearrangement allows the second terms

in these expressions to be treated explicitly as a deferred source term, thus
enabling the strict tri-diagonal nature of the discretized equations to be
retained in implementation.
It will be seen from the above discussion that SOU solutions in one
dimension are bounded, stable and second order accurate. Extension to multi
dimensions is straightforward and is based on individual one-dimensional solutions, characteristic of profile schemes, Section (6.1). However, in multi
dimensions due to possibility of generating negative coefficients, albeit
minor, see equation (6.28), the solutions will not necessarily be bounded.
Furthermore, the above one-dimensional analysis carried out for uniform sign
of the convecting velocity does not strictly ensure stability for velocity
changing sign across a cell, characteristic of recirculation zones. Use of
Second Order Upwind differencing by Coakley (72) encountered oscillations in
the vicinity of a shock where the Q eigenvalue changes sign across a P-shock.

-79-

Finally, characteristic of profile schemes, some smearing of profiles is'

expected, as the scheme does not explicitly account for the misalignment
between the grid and the streamline.
6.2.2 Solution Details

The matrix equation generated by the finite difference molecule implied

by (6.28) and its two-dimensional analog is not generally amenable to solution
by direct application of the Tri-Diagonal Matrix Algorithm (TDMA) or its
Alternating Direction Line Gauss-Sidel (ADI) equivalent; variants of Penta(30)
Diagonal Matrix Solvers formulated in ADI fashion
have to be used with
attendant computational costs. Alternatively, discretization relations provided by equation (6.35) that retain the strict tri-diagonal nature of the
formulation could be used (adopted in this study). However, iterative stability of such solutions incorporating a deferred correction strategy is not
assured a priori in multi dimensions and solution costs might be relatively
high.
A further point in obtaining solutions using the SOU scheme concerns the
application of boundary conditions. Due to the extended nature of the molecule, near the boundaries the scheme must be modified using conventional practices, including the use of Central or lower order Upwind differences. Shyy
(73)
et al
examined in detail the influence on the solution of adopting
implicit or explicit treatment of boundary conditions at the outlet boundary.
(29)
It is also claimed by Gupta and Manohar
that the overall accuracy of the
scheme is severely affected by the inconsistencies introduced in boundary condition specification. In this study Hybrid differences were used near "flow
boundaries" on the grounds of robustness.
6.3 ADVANCED SKEWED UPSTREAM DIFFERENCING SCHEMES (SUDS)
6.3.1

Influence Point Equations

The two SUOS variants considered in this study are primarily designed to
reduce the error associated with the evaluation of convective fluxes at the
interfaces, C in equation (6.5). In both schemes this 1s accomplished by
i) breaking each control volume face into two parts thereby effectively increasing resolution and, ii) improving the evaluation of the influence of
nodal values of 0 for control volume evaluations.

-80-

To increase the resolution of the convective flux evaluation, each control volume face 1s broken up into two parts. For example, the e control volume face is broken up into the parts shown in figure (6.2). As a result, the
convective flux evaluation at the e face is given by:
(Pu0)edy Ay

where

(6.36)
<*

ene

Considering, now the evaluation of J

and introducing the previously used
approximations; i.e.,
i) (pu)e is constant along the ese portion of the e face,
ii) 0 varies at most linearly along the ese face,
then
J

ese * (pu)ese *ese

(cf. equation (6.5))

(6.37)

where 6ese is the value of 0 midway along the ese face.

It is therefore necessary to evaluate 0
in terms of the surrounding
ese
P,E,S and SE nodal values of 0. The approach adopted here is to ensure that
the evaluation of 0
is based on the physical influence of the nodal val656
ues of 0 on 0QCO, Stubley.v ' For this .reason, 0 65 C is referred to as
an influence point value of 0.
The next step in the discretization is to define how the influence point
value of 0 is related to nodal point values. However, before proceeding with
this, it is necessary to introduce some additional terminology. First, borrowing from the finite element approach. Section (5.1.6), the four nodes surrounding the influence point can, in fact, be seen to define the four corners
of a rectangular element, and the element, so defined, is made up from quadrants of four different control volumes. Also, by introducing two influence
points along each control volume face in domain, each element contains four
influence points. The configuration for the element containing 0
is
shown in figure (6.3).
-81-

nne

nnw

<P ene

'wnw

ne PORTION
OF e-FACE

>

-w
I

wsw

ese

se PORTION
OF e-FACE

sse

Figure 6.2

Division of Control Volume Faces

NODAL POINTS

INFLUENCE POINTS
CONTROL VOLUME FACE
ELEMENT BOUNDARY
CONTROL VOLUME QUADRANT

Figure 6.3

Element Surrounding Influence Point Value 0ese

r~
ii
i
r*n_b.

ese

sse

-X-

i
SE

N1679

Figure 6.4

Evaluation of 0U for LP-SUDS

-82-

It now remains that 0ese must be expressed 1n terms of nodal values of

0 and that the influence of the nodal values of 0 on 0ese must reflect the
correct physical influence. As suggested by RavrMO)' one way of accomplishing this is by way of an approximate algebraic representation of the relevant
differential equation. This is discussed next to derive the Influence point
equations for the linear profile variant of SUDS, (LP-SUDS).
The LP-SUDS influence point equation for face values of 0 can be arrived
at by considering equation (6.1) at each influence point. For instance, at
the ese influence point, equation (6.1) can be written as:
C

ese

= D

ese

+ S

(6.38)

ese

where C represents the convective terms, D, the diffusive terms and S the
source term. To derive an influence point equation, an approximate representation of each of the terms is required for each influence point.
Consider the convection terms, C, written in non-conservative Cartesian
form,
C =

30
30
V
'" 37 + " 3^=

PV

.. 30
K

where V = \ u2 + V2 and ds = u/V dx + v/V dy

To approximate the above equation algebraically, original skewed upwinding ideas are upwinding employed such that

-0
C

ese

= pV ()

pV

<

U)

(6 39)

'

where 0 is an upwind value of 0 and L is the distance between the upwinded

point and the influence point.
To illustrate, consider the case shown in figure (6.4) where the streamline through 0ese is such that it intersects the s control volume face to
the left of 0 .In this case the value of L is as indicated and the
ese
upwind value of 0 is linearly interpolated between 0
and 0
sse
s

-83-

b "sse

+ (1

- b> *s

' ., - - <6'40)

where 0
1s in turn linearly interpolated between the nodal values of 0p
and 0 . The general rule for the calculation of 0 and the length scale L
is to take the local streamline, passing through 0ese , upwind until it
intersects an edge of a control volume quadrant. Then a linear interpolation
is used along that edge between appropriate nodal values or influence point
values. Therefore, if the upwind portion of the streamline intersects a control volume quadrant edge between two influence points, the value of 0 is
determined by linearly interpolating the two influence points. Similarly, if
the upwinded streamline intersects the line joining two nodes, the linear
interpolation of 0 between the two nodes is used. Implementations using variants of such practices are discussed later in Section (6.3.3).
As discussed above the value of 0 along quadrant edges is determined
by linearly interpolating between nodal or influence point values of 0.
Equivalently, 0 is determined by assuming a linear profile of 0 between
nodal or influence points. Hence, the resulting scheme 1s referred to as linear profile SUDS. In a later section a different evaluation of 0 adopted
in Mass Weighted SUOS is described.
Considering next the evaluation of diffusion terms in equation (6.1),
i.e.,

a.(U M) + a_,
at,
ax 0 ax' ayU0 ay'
means have to be provided to ensure that no non-physical Influence results due
to their incorporation in the influence point equations. Thus D is usually
split into two parts, the component normal to the control volume in question,
D , and the tangential component, D . For 0 , these are:

a U(r
. 30.
;
0 = ax
0 ax)
nn

-84-

An approximate representation of Dn is given by:

Dn

(0_ - 0
)_i
ese _ 0
ese - POf

f r
'

AX+

AX"

)..

(6.41)
(AX* + AX~)/2

where the distances AX and AX are indicated in figure (6.5) and f is a

correction factor. The correction factor f, is required for two reasons.
First, if upwinding is used to represent the convection terms at the influence
point and Central differencing is used for 0 , then after combining C and
D , the most accurate evaluations of 0
require that f range from 0 to 2
depending on the grid Peclet number. Secondly, the correction factor is required to ensure that negative downstream coefficients, which are non-physical
and can result in spurious spatial oscillations, do not arise in the control
volume equations. It can be shown that for this condition to be met it is
sufficient to assign f the value of 0.5. In the interest of simplicity this
assignment is adopted in the present study.
For the tangential component of D, the following form is adopted:

D* a T

t -0
e ese

Ay + /2
(6.42)
(A/ + Ay~)/2

where 0 and 0
and distances Ay and Ay are as shown in figure
(6.5). To evaluate 06 -and 0 oC in terms of nodal values linear interpolation can be used. However, to ensure that no significant negative downstream
coefficien" arise, the values of 06 and 0 O C are determined from the
coefficients
relations:
0e = (1 - A) 0p + A0E

(6.43)

0$E = (1 - A) 0$ + A0S

(6.44)

-85-

T
1
1
1

1
1
1
1
ft .

Ay*
fxgse

i'
t

1
|

::
$ se
A Y~~

SE

Ay-

I
1

Figure 6.5

|:

AY_^

Evaluation of Diffusion Influence for SUDS Schemes

I
1
\

mese

2. \

msse
m^

ll

_T7

^ ^v
s'
^ese

f3

^r

it* '

,>

1S>Ei

s
N1680

Figure 6.6

Evaluation of 0U for MW-SUDS

-86-

where
wPe
Pe
A = (1 - e )/(! - e )

(Ax+ + Ax

Pe = pu

w = ax /(AX + AX~)

which can be derived by assuming that transport of 0 is locally onedimensional along the x-coordinate direction.
To arrive at the final form of the influence point equation consideration has to be given to approximating the source term, Sese If no source
term influence is desired, then S =o. However, to include this source
ese
term influence (STI), then, the evaluation of S
is based on the known
ese
values of S in each of the surrounding control volumes. In the present study
if the x component of velocity at 0
is positive, then S
is taken to
be the value of S in the P control volume. Otherwise, S
assumes the
value of S in the E control volume.
6.3.2

Assembly of Influences and Discussion of the Finite Difference

Equations

Combining together the above approximate representations of each term of

the influence point equation, i.e.,

ese

= n
4- n^ 4. ^
"ese uese ^ese

(6.45)

provides a relationship between 0 , the four nodal values of 0 as well as

the three other influence point values of 0 contained in the element, figure
(6.3). It is important to re-iterate, that, as discussed by Raw,
it is
essential that equation (6.45) adequately accounts for the appropriate
physical influences.

-87-

Having described the manner in which the influence point equation for
0
is derived, it is a simple matter of repeating the procedure for each
ese
influence point contained in the element. The result is a set of four linear
algebraic equations at each influence point involving four nodal values of 0
and four influence point values of 0. It is then a simple matter of Inverting
a 4x4 matrix to be able to express each of the influence point values of 0 in
terms of the four nodal values of 0, Section (6.3.4). Moving from element to
element, then the same procedure can be repeated so that each influence point
value of 0 in the computation domain can be related to surrounding nodal values. Finally, substituting these relations for influence points into the
evaluations of convective fluxes i.e., equation (6.5), the control volume balance between convective fluxes, diffusion flux and source terms given by:
C
C
C
C
C
C
C
C
D
D
Jese
+ Jene
- Jwsw
- Jwnw
+ Jnne
+ Jnnw
- Jsse
- Jssw
= J
e - Jw - J
n - Js + S0

can be expressed in terms of nodal values as:

VP

VE

Vw

VN

Vs + 'NE'NE + VNW + aSE*SE + VSE * bp

(6.46)

Along with the representation of boundary conditions, the result is a set of

linear algebraic equations for which a solution for nodal values of 0 can be
found.
As will be seen from an examination of (6.46) nine nodal values of 0 are
involved. The coefficients of equation (6.46), a_,
a_u
a_OWu are such
r,
that
a > a + a + a + a + ' a
+a
+ a
+a
P E
W
N
S
NE
SE
NW
SW

Also, by the specific choice of algebraic relations used to represent the

influence point convection and diffusion terms, the coefficients on the nodal
values downstream of 0 are not negative, thereby reducing the occurrence of

non-physical oscillations in the solution. However, a rigorous stability,

accuracy and convergence analysis of the kind described in Section (6.2.1) for
the SOU scheme is exceedingly difficult to accomplish for equation (6.46)
a priori, and hence such characteristics will have to be established by computational experiments, Section (7.3.2).
A close study of equation (6.46) further reveals that, because a linear
profile assumption is made to express 0 in the influence point equations in
terms of the surrounding values of 0, the coefficients of nodal values that
are not directly downstream of 0 may be negative. To illustrate, consider
uniform flow angled to the grid Ax = 0.1, ay =0.1 such that u = 0.5, v = 1.0
and Y - 10 . For this case the a coefficients normalized by BL are:
3

NW 3N 3NE
3

-asw

PaE

0.0

0.0

0.0

-0.1153

1.0

-0.1154

.0.5769

0.6923 -0.385.

As a result there exists the possibility of spurious oscillations in the solution of 0. Although these negative a coefficients do not produce any large
oscillations as will be demonstrated in Section (7.3.2), there are some applications such as the k equation of ^ - e turbulence model, where it is
essential that the solution be bounded. In this instance the Mass Weighted
SUDS which is described next, can be used.
6.3.3

Mass Weighted Skewed Upstream Differencing Scheme (HW-SUDS) for

Positive Definite Variables

As discussed by Raw,
to prevent negative a coefficients it is necessary to prevent any node external to the control volume from having a net
effect of transporting a convected variable out of the control volume. To
accomplish this an alternative evaluation of 0 in equation (6.40) is required where now, 0 is determined by logically deducing where the mass
crossing the ese face originated. Consider the case shown 1n figure (6.6)
where, m , the mass flux through the ese face,is positive and known. The
ese
first step is to assume that there are two possible contributions to m
ese
1) from m , the mass flux through the sse face and 11) m , the mass flux
through the surfaces formed by joining the locations 0 , 0 , and 0 .
The second step is to determine m from mass conservation, I.e.,

-89-

- msse
Now, if m > m , then all the mass crossing the ese face originated
sse ese
from the sse face. Therefore,
V'sse'

(6 48)

'

By similar arguments,' if m p >- mese then,

*u = *P

(6>49)

Finally, if m
> m
and m
> ITL, then both m
and mn conese
sse
ese
P
sse
P
tribute to m
in the proportions
ese

m
~^
m
ese

m
and m-
ese

, respectively.

Therefore,

'&*%''

(6 50)

'

Combining all three cases together, the evaluation of 0 in equation (6.49)

is replaced by:

"u = "sse *(1"a) *P

(6

'

where a = MAX (MIN (m /m , 1.0), 0.0). Considering the evaluation of

o given above, the resulting discretization is appropriately referred to as
Mass Weighted SUDS (MW-SUDS).
Adopting MW-SUDS, the a coefficients normalized by a for the same
uniform angled flow case described previously becomes:

-90-

NW

0.0

0.0

0.0

0.0

1.0

0.0

.0.5

sw

0;5

0.0.

It is evident that mass weighting has removed

the occurrence
of negative a
*
coefficients but, although it is not obvious from the coefficients, some false
diffusion has been introduced. However, the resulting errors are considerably
smaller than those that arise when the Hybrid 5 point discretization schemes
are employed. Illustrations of this point are made in Section (7.3.2).
1

6.3.4

Further Considerations in the Implementation of Advanced SUDS Schemes

The LP-SUOS and MW-SUDS schemes considered above, involve evaluations of

the coefficients of influence point equations that require the inversion of a
4x4 matrix in each flux element. These costly inversions are required because
each influence point is implicity related to adjacent Influence points. In
this section an alternate explicit evaluation of the coefficients of the influence point equations is described which does not require the inversion of a
4x4 matrix.
Considering equations (6.38) and (6.39), an explicit formulation for
LP-SUDS is accomplished by computing 0 and length scale I in equation
(6.39) in a manner compatible with upwinding the local streamline passing
through A ese until it intersects an edge of the flux element, figure (6.4).
Then, a linear interpolation is used along the flux element edge between
appropriate nodal values. Upwinding to flux element edges instead of to control volume quadrant edges as in the implicit approach is necessary to ensure
an explicit influence point equation.
To evaluate the diffusive term of equation (6.38) at the ese influence
point, D
is determined from a bilinear interpolation of the approximate
values of D available in each of the surrounding control volumes. Note that
in the interest of simplicity best estimates of the nodal values of the variables are used to evaluate the control volume D's. Naturally, this deferred
correction approach requires iteration which is readily incorporated into the
iteration procedures used to solve flow problems.

-91-

The evaluation of the source term Sese for the explicit influence
point equation is determined in a manner similar to D , from a bilinear
interpolation of the surrounding nodal values for S.
The explicit influence point equation for MW-SUDS is very similar to the
implicit approach. As in the implicit approach, the upwind value of 0 is
based on a mass weighted average of surrounding nodal and influence point values of 0. In the explicit approach developed by Huget,
0 is given by
a similar but not identical relation to equation (6.51), i.e.,

0u = <*0'
+ (l-o) 0_
sse
P

.,

rov
(6.52)

where 0' is not equal to 0 . Instead, 0' is given by:

sse
sse
sse

"sse ' B'sese

where B = MAX (MIN (-msese /msse , 1.0), 0.0) and m sese is the mass flow
crossing the quadrant face directly south of the ese face. To ensure that no
negative a coefficients arise,

>P

(6 54)

'

For the explicit MW-SUDS the diffusion and source terms of the influence
point equation are as described for the explicit LP-SUDS.
Finally, the evaluation of 0
as determined above is appropriate for
high grid Peclet numbers. However, for low grid Peclet numbers a bilinear
interpolation is more appropriate and for intermediate grid Peclet numbers a
combination of the two evaluations can be used, exploiting the flux blending
strategy discussed in Section (5.15). Specifically, the Skewed Upstream differencing representation and bilinear representation of 0 , respectively
h
ese
denoted by 0
and *ese. can be combined in the following manner:

-92-

*ese ' ' 4e +(1'v)

*ese

<6'55>

where

Pe2
5+Pe2
The given evaluation of is chosen so that the introduction of a bilinear

interpolation for 0ese does not introduce any negative a coefficients. Note

I I
s
that the above equation
forL.0cse
= ***<>
a ensures that for iPe i 1 , '_

ese
ese
and for |Pe|

6.3.5

1, 0 e

= 0 ese'

Boundary Conditions

Consistent treatment of the boundary conditions compatible with formulation of LP-SUDS and MW-SUDS described in the previous sections, has been
implemented by Raw (10) and Huget (30) for scalar as well as flow equations. However, in this section a specific treatment of boundary conditions
appropriate for use in the Pratt and Whitney 2-D TEACH code will be considered.
The default layout of the numerical grid adopted in most production
codes based on TEACH structure is such that, for a scalar, the boundary is
straddled by nodes. However, the Pratt and Whitney 2-D TEACH code (Appendix A
describes a similarly structured 3-D TEACH code) overrides this default by
moving all fictitious nodes that are outside the computational domain so that
they are coincident with the boundary. In the flux element structure of
LP-SUDS and MW-SUDS this shift of fictitious nodes is readily implemented and
results in flux element boundaries, quadrant boundaries and the boundary of
the domain being coincident, figure (6.7). Note that no adjustment of the
flux element is required for u and v velocity nodes which, by virtue of the
staggered grid, naturally fall on the boundary. This includes u velocity
nodes along east and west boundaries of the domain and v velocity nodes along
north and south boundaries of the domain.
In addition to the adjustment of the grid layout in the vicinity of
boundaries, the Pratt and Whitney 2-D TEACH code also assumes that for control
volumes adjacent to the boundaries, alit. , aNW ajt and a on are zero.
Since this assumption permeates the entire code, any change regarding this
assumption would require a complete restructuring of the code. Instead, the

-93-

- ^ B O U NDftRY OF DOMfllN
*r

/
/

\:
*

X ELEMENT BOUNDflRT
1 FLU
/

' OUfiORflNT
^BOUNDflRT

A^"

INTEGRflTION

>
*
\
/

/NODE POINT

FI CTITIOUS ^
NODE
^

>i

(a)

Figure 6.7

>

//

[-* -H

\*^^

/ COINCIDE NT
/ /BOUNDfiRI ES

(b)

Default Flux Element Layout and Adjusted Layout 1n Vicinity of

Boundary

-94-

assumption is retained and only MW-SUDS is used in flux elements that are
adjacent to boundaries. It can be shown that the use of MW-SUDS will generate
coefficients which satisfy this assumption.
Finally, for axlsymmetric problems, all length, areas and volumes of a
flux element are calculated accordingly, and any additional terms that arise
in the differential equations representing momentum conservation are included
in the integration point source term.
6.4

COMPACT IMPLICIT DISCRETIZATION SCHEMES

Most of the discussion pertaining to Compact Implicit discretization

schemes will be developed in a one-dimensional context for clarity.
Extensions to multi-dimensions are given in Section (6.4.1).
6.4.1

Derivative Compact Implicit Scheme (PCI)

Consider the Cartesian, linear convection-diffusion equation of (6.1)

with p=l ,
L x = S(x)

(6.56)

where

is the spatial differential operator with spatially constant diffusion coefficient 0 and convective coefficient u and the source term S is a function of
the independent variable x.
The DCI approach is most readily described by first considering the discretization of equation (6.56) using standard second order centered difference
representations of the derivatives, (uniform grid of spacing h):

0.

- 20, + 0, ,

J1_J

in

h'

-95-

+0(h2}

(6>58)

where 0. is a discrete approximate value of 0 at x=x. and h is the spacing

between discrete points. Using equations (6.57) and (6.58) in (6.56), an
explicit relationship can be derived that governs the evolution of 0 in the
form:

(1 + IT) 01+1

2
= r S.

- 201 + (1 - Y~) 0 i _ 1

(6.59)

where Pe is the grid Peclet number.

The DCI procedures depart from the usual explicit form of discretization
described above by seeking relations between the derivatives in equation
(6.56) and 0 that implicitly involve evaluations of the derivatives at neighbouring discrete points, Section (6.1.1). DCI schemes were discussed by
Collatz,
Krause et al
who referred to their use as Mehrstellen
integration, Adams
under the name of Hermetian finite differences and by
Kreiss as reported in (31), who coined the Compact Implicit terminology. The
standard DCI approximations to the derivatives of equation (6.56), given by:

40.

0J.1

1+1

0.J.1

J+l

~ Q*

1-1

, n^x

c.n\
(6.60)

/c

0(h >)

(6.61)

where 0' and 0" are representations of d*/dx and d 2*/dx2 respectively,
at discrete locations, can be derived by simple Taylor series analysis.
Because of the implicit nature of equations (6.60) and (6.61), the simultaneous solution of the following equations is required:

t "
'
Vl + U*1= Si

(6.62a)
= J (01+]

-96-

- 0^.,)

(6.62b)

- 2t)i + "W

<6'62c>

for the values of 0, 0', and 0". Equations (6.62) are easily solved by the
(79)
application of a block 3x3 Tr1-D1agonal Matrix Algorithm.
Incorporation of boundary conditions in (6.62) is accomplished via
either fourth order one-sided representation of the various derivatives using
Taylor series or Fade' approximations. Rubin
Rubii and Graves
derived the
following fourth order representation of 0'.
- hi (20J+1 + J) =0 + 0 (h4)
6

(6.63)

Alternatively, a fifth order Pade1 approximation relating the first and second
derivatives is given by:

"

2 (01 *

The DCI method discussed above is formally fourth order and can be cast
into a block tri-diagonal form. -The major shortcomings of DCI are:
1) the inherent grid Peclet number limitation, restricting Pe to be
of order 2
11) the need for closure equations on the derivatives at the boundaries
ill) the need to solve sets of block tri-diagonal equations with particularly large blocks (7x7 for a scalar in three dimensions).
There are several ways to overcome these drawbacks. One approach is
given by the subclass of Compact Implicit methods called Operator Compact
Implicit (OCI) methods. Several OCI schemes are considered in the following
section.
6.4.2

Generalized OCI Schemes

The classical OCI method, first postulated by Swartz(81) and later by

(36)
Ciment et al,
assumes the existence of tri-diagonal relationship between

-97-

f\ v

0 and the discrete representations of L denoted by L

forms of equations (6.57) and (6.58), i.e.,

, guided by the

or

R and Q are tri-diagonal matrix operators. The coefficient of R and Q can be

determined in a number of ways including the Taylor series analysis provided
in Appendix B.
COCI adopts the following tri-diagonal operators R and Q to discretize
equation (6.56) in the form given by (6.56):
qP = 60 + 16 (uul - u^z

-4uui u^z2

q6 = 6 + (5ui - 2u.M)z
qw = 6 - (5u. - 2u.+1)z

(6.66)

r6 = -^ [qW(2-zu.
) * qP(2+zu.) * qe(2+3zu.
. j]
n
1+l
^
'
^
= ? [qW(2-3zu. ,) -i- qP(2-zu.)

e +r
w. ) , z =h
rP = - /( r

An alternative derivation, valid for variable grid spacing is presented

by Ciment, et al.(36)
It will be seen from the above discussion (equations (6.65) and (6.66))
that for one-dimensional problems, COCI solutions can be obtained using a
A X

single scalar tri-diagonal decomposition when*

* = LX = S(x)

-98-

is known from
equation

(6.56)

Further, COCI does not require boundary closures for higher order derivax
tives. Instead, evaluations of L at boundaries is required, but these are
readily determined from equation (6.56) and the supplied conditions of 0.
Therefore, by adopting the OCI approach two of the three drawbacks of DCI have
been overcome. However, there still remains the problem of a restrictive grid
Peclet number limitation. For the COCI scheme derived in Appendix B
analysis
has shown that solutions remain bounded for Pe <, ^/FF = 4.208.
However, the behavior of the numerical solution can be very non-physical at a
considerably lower value of Pe, Section (7.3.3). In fact, non-physical
behavior results when Pe>2. In attempting to remove the restrictive Peclet
(37)
number criteria, Berger, et al
were led to the development of a Generalized OCI (GOCI) family of schemes which suffer from no such restrictions.
These will be discussed next.
Berger, et al began by mathematically summarizing the properties of the
tri-diagonal matrix operators R and Q of equation (6.65) that are required to
ensure a physical (i.e., bounded) solution to
/ QO \

0 (0) = a ; o(l) = a ; r>0 ; u>0 (i.e. advection from x=l to

u
I
0
x*0) and S(x) >0.
These properties are postulated to be:
w
i) r >0 (downwind influence of 0 cannot be negative)
g
ii) r >0 (upwind influence of 0 must be positive)
p
e w
iii) -r > (r +r ) (diagonal dominance)
iv) qe>0, qw> 0,q > 0 (operator influence must be positive)
e w
v) r >r (upwind influence on 0 must be as large as downwind
influence on 0)
p
e
w
vi) q u.> q u. , + q^ (to ensure invertability of Q)
A review of the coefficients of R and 0 of COCI, equations (6.66) reveals that the above requirements cannot be satisfied for all grid Peclet num
bers, thus it is necessary to re-examine the structure of truncation error
expression given in Appendix C.

-99-

The truncation error, T, associated with equation (6.65), i.e.,

- 0 (L*)

(6.67)

where *. and Lix are defined as the exact values of * and Lx respectively, at the i'th discrete location, can be expanded in a Taylor series,
r. T*i + T1*] + T2** + T3*3 + T4$4 + .

(6.68)

where * = d */dx , see Appendix C for definition of the T terms.

3
4
Unlike the COCI method that strictly assumes T = T =0, a family of
possible fourth order OCI schemes results by relaxing the constraints on
T3 and4 T such that, T3 = 0 and T =4 0(h4 ), Unfortunately, as shown
by Berger et at (37) , there is still no unique evaluation of the
coefficients of R and Q which satisfies the above requirements. However, by
further relaxing the constraints such that
T3 = 0(h4) ; T4 = 0(h4)

(6.69)

then, a family of fourth order OCI schemes results that satisfy the above
requirements.
i
By recognizing from equations (6.66) the R can be expressed in terms of
Q, the family of OCI schemes resulting from equation (6.69) can equivalently
be viewed as a family of coefficients for the Q matrix operator. Furthermore,
it is postulated that the coefficients of Q can be expressed as the following
polynominal functions of z (cf. equation (6.66)):
w
w,0
q = q +

q
q

w,l
w,2 2
q z + q z +

w,3 3
q z

w,v
+ q z

(6.70a)

e,0
e,l
e,2 2
e,3 3
= q
+ q
z + q
z + q
z
+

+ q

e,v v
z

(6.70b)

p,0
p,l
p,2 2
p,3 3
= qK
+ q
z + q
z + q
z

+ q

p,w v
z

(6.70c)

where the coefficients of the polynominals are functions of u and are independent of r 0 (Note that the 0 of COCI belongs to this set).
From equations (6.69), the evaluations of T given in equations (C.4) in
Appendix C and a substitution of the polynominal functions for Q gives:
-100-

T 3 = rh
0
T

h2
= ^-

[t 3 ' 0 + t 3 > 1 z + t 3 ' 2 z 2 + 0(z 3 )] = 0(h 4 )

(6.71a)

[t4'+ t 4 J z + 0 ( z 2 ) ] = 0(h 4 )

(6.71b)

where
t 3 ' 0 = qw'-qe' - 0 ( h 3 ) f t 4 ' 0 = qP' - 5 (q e ' + qw') = 0(h 2 )
,.3,1
t

4,1

w,l e,l 1 .. e,0

= q
-q
(2q
U

p,l . . w,l

= qK> - 5 (q

t *

'

W 2

=q ' - q ' -

. e,0

e,l.

^ q , ) - (q
'

6 2

. w,0
+ 2q U_

(2q

w,0

u^..

eJ

p,0 .
... 2.
- q K> u = 0(h )

....

u^-,) = 0(h)

'

W l

'

'

''

+ 2q ' u1_r q"' ^) = 0(h)

Further analysis of equations (6.70) with respect to the requirements

i
(i) to (vi) reveals that the following conditions on the highest order (v=m)
polynominal coefficient must be satisfied:
q

e.m > 0- , qw,m = ^

P,m
e.m
0 qHt U = q U

., -.^^

(6.72)

and that =3 in equation (6.70) is the lowest order of the polynomials.

Based on the results of equation (6.72), the coefficients of equations
(6.70) can be selected. Begin by choosing q =60 consistent with COCI
results. Setting t3' = t4' = 0 it then follows from equation (6.71) that
q

e,0= qw,0 = ,6

.,
_..
(6.73)

Making use of the fact that,

1/2 (u.

+0(h)
+ 0(h)
* u^) = u. + 0(h)

combined with the t evaluations of equations (6.71) it can be shown that,

-101-

P>1

ui 1

(6.74a)

" 3ui * 0(h

0(h2)
w,2

' =

(6.74b)

u.
Pt

* 0(h)

(6.74c)

Finally, from equation (6.72), it is concluded that,

qW'3 = 0

(6.75a)
e,3

A 1

The ,remaining unknown coefficients in equations (6.70) are q

W>2

,q

f\ 2

6 3

q
and q ' . Substituting equations (6.73), (6.74) and (6.75) into
equations ;(6.70) and simplifying gives:
w
,
q =6 + |[P1 - 3)Pe + p 2 Pe

(6.76a)

= 6 + (;p1 + 3)Pe + (p1 + p 2 )Pe

P _' tn s.

+ P4Pe

(6.76b)

/t

where the yet unknown constants are:

P1 = q Ptl /10u i

; p2 = q W ' 2 /u 2

qP'2/u2 ;

q6'3/u3

The choice of any one of p, to p. is arbitrary with the remaining

1

p's to be evaluated from requirements (1) to (vi). One choice is to set

p =3. For uniform u (u.=u =u ) and p =3 it can then be shown
that,
-102-

Tt-\

p =0

(6.77a)

P3 = 9

(6.775)

P4 = 1.5

(6.77C)

Substitution of equations (6.77) into equations (6.76) leads to a specific GOCI scheme valid for generating bounded solutions for all values of Pe
with u constant, which numerical results have tended to suggest is optimal
qW = 6

(6.78a)

qe = 6 + 6Pe + 3Pe2 + l.SPe3

(6.78b)

qp = 60 -i- 30Pe + 9Pe2 + l.SPe3

(6.78c)

and from equations (6.66)

rw = 72

(6 78d)

'

re = 72 + 72Pe + 36Pe2 + 12Pe3 + 3Pe4

rp=-(rw+re)

(6.78e)
(6 78f)

'

It is important to note that the r coefficients given in equations

(6.78) are such that r /r = e + 0(h ) and it is this exponential
character which is, in part, responsible for ensuring that the resulting solution is bounded.
Error analysis by Berger, et al (37) demonstrates that GOCI schemes
such as the one given in equations (6.78) are formally fourth order (i.e, in
the low Pe limit) but automatically change their order of grid convergence as
the grid Peclet number increases. In fact, in the high Peclet number limit,
GOCI method is only second order accurate and at intermediate values of Pe the
order may drop to zero.
In summary GOCI scheme is, for one-dimensional problems, bounded for all
values of Pe. In addition, GOCI does not require additional boundary equations for the operator, and solutions to the algebraic equations are readily
obtained using TDMA for a single scalar. Variations on the above GOCI
approach which sacrifice formal fourth order convergence for uniform second
order convergence exist. One such approach will be used as the basis for a
new control volume based OCI scheme presented next.

-103-

6.4.3

Control Volume Based Operator Compact Implicit Method of Exponential

Type (CVOCI)

In this Section 1s developed a particular scheme for the family of exponential OCI schemes that ensures conservation 0 over discrete control volumes. Most, if not all, existing Compact Implicit schemes reported 1n the
literature are of a "point finite difference" nature and hence do not strictly
satisfy some of the desired attributes of a discretization scheme discussed in
Section (4.2).
Considering equation (6.56) with spatially varying diffusion and convection coefficients, i.e.,
L

* = d7(ra df>+ dx"

(U4>) = s(x)

equation

(6.56)

a finite volume integration can be performed over the length of control volume ,h, to yield
:

h
x

'Jef -Jwf -/s(x)dk

equation (6.13)

where 0c I and JW I denote the interface fluxes of 0 given by

;J

wf

To obtain a discrete algebraic representation of equation (6.12) in

terms of 0 and , the discrete approximations of * and L are required
x
within the control volumes; it is assumed that both L and S(x) vary at most
linearly over the control volume, thus,
u

ii

f Lx dx * hx

-104-

(6.79a)

f S(x) dx =
0

(6.79b)

Also

Ci

(6.79C)

= S.

"

In equations (6.79) jC * and S., are approximations of L and S(x) at

the midpoint of the control volume. The approximate representation of the

flux expressions in equation (6.13) is not so simple and requires evaluations
of d*/dx and * in terms of 0 and L at both control volume boundaries.
In the present exponential scheme, these evaluations are determined from the
analytic solution of equation (6.56), i.e.,

udj

where 0<x<h, *(x=0)= 0rt, *(x=h)=0

and r and u are assumed
to be
1 0
.
.
constant over the range of x and L is an arbitrary function of x. The
solution of the above equation, with no further approximation is given by:

11
1
(i -) r^ "

i/- .

(F(X)

"

F(h)

V1

(6 80)

'

where
F(x) =

P - - Re - -

LX dx - eZ

Lxdx]

j;x = uxr

uh
0

The gradient of 0 is easily shown to be:

(G(x) + F(h)

-105-

(6.81)

where
Lxdx

G(x) = - e

Equations (6.80) and (6.81) require the Integrals of L 1n the evaluation of

x
x
F and G. Unfortunately L is not known. Therefore, profiles of L have
x
to be assumed. A particular choice that sets L to zero yields the wellknown Exponential Differencing Scheme, (EDS). (83) With this in mind, for
x
'

any non-zero L profile assumption, equations (6.80) and (6.81) can be seen
to provide the corrections to EDS necessary to account for the influence of
x
L = S(x) on the solution for A. Since in this OCI approach the operator 1s
x
implicit, alternative profiles of L can be considered.
x
Assuming that L varies linearly with x, equations (6.80) and (6.81)
become:
V

f
4(x) = 0 + (0-0)[ ] + ~- (x-h
e -1
e -1
S:iL

2
r x
e
aa
h2 f*9h e\ Zf-l
u (/x_
/n_
'-\\
(
)[
h
- 2 ~ u " 2 ' u P ,])

/c ooa\
'

(6 82a)

( uu+ hepP-l

Xx -j r x

--

r2

2 rh

Substituting equations (6.82) Into the flux expression of equation (6.13), the
e face flux becomes

ef

O+

"l+

0+

-106-

jC

(6 83a)

'

where

r(* PeP
~w
r* = -* *f-

(6.83b)

e -1
^-"iTp*
" e -1

(6.830

q" = h(* - * - - ! + ir-f J

J)
2tT P P
2(e -1) P(e -1)

(6.83d)

_
2
e
q = h(*-r - ^ + Lp
1 )
2h^ p*
2(e -1) P(e -1)

(6.83e)

Noting that 0Q=0. 1=1+1. X J = JC^ and ^= *-^t equation (6.83a)

can equivalently be written as:

ef

=r

*1* r *

+q

+q

(6.83f)

Similarly, the flux through the w face is given by:

Jwf = ^V-, + rV. + q" *_., * q e *

(6.84)

Combining equations (6.79), (6.83) and (6.84), equation (6.13), the control
volume flux balance equation is represented by:
-x + qe ..,
x
rw 0. , + rp0. + re0. , = qw-x
J,._, + qp JL.

where

w
r
e
r
rPK
qw

-w
= -r
_e
= r
/ w +r
= -(r
= ~w
q

e
-e
q = -q

u e-q
qHP = h-q

-107-

,,
OCi
(6.85)

Finally using (6.79c) to evaluate -C the resulting profiles for 0 can

be determined from the solution of equation (6.85).
At this point it is useful to review some of the characteristics of this
discretization scheme. The scheme is an OCI variant of exponential type and
is conservative. Further examination of the coefficients of equation (6.85)
also reveals that conditions (i), (ii), (iii), (v) and (vi), Section (6.4.2),
required to ensure bounded one-dimensional solutions are satisfied. However,
it can be shown that condition (iv)is not always satisfied. Consider,
e e
2
for example, the limiting case P = where q =-q = -x /(2h) < 0
e
which is in violation of condition (iv). Physically, a negative value of q
implies that if the value of L . . which equals the value S. , is
increased, the value of 0. decreases when in fact the value of t should
increase. The result is a one-dimensional solution for 0 which is not neces_g
sarily bounded. To remedy this situation, the evaluation of q must be mod_g
ified so that it remains negative. Fortunately, positive q values arise
/\e
-e
only for large P values. One way to obtain q , a corrected q value, is
from

- MAX (qe, 0.0)

(6 86a)

'

Then to ensure that condition (vi) remains satisfied, it is necessary to also

modify q .

- MAX(qe, 0.0)

(6 86b)

'

By similar arguments, it can be shown that when P is large in magnitude

but negative, ~q can also become negative. In this instance q and q
are given by:

- "IN (?. 0.0)

= q6 - MIN (?. 0.0)

-108-

(6 87a

'

>

<6.87b)

Finally, 1n terms of q and q , the q evaluations of equation (6.85)

which satisfy all conditions for a bounded one-dimensional solution are given
by:
qw = q

q = -q
p .
qK = h - q -q

Adopting the correction strategy outlined above, it is observed that

corrections to q are required only when the downstream value of q value
becomes negative. The effect of applying the correction, when required, is to
make the downstream of q value equal to zero and increase the upstream q value. The resulting q values are then equivalent to those which would have
arisen from assuming a constant value of L(x) in equations (6.80) and (6.81)
with the upstream value of L(x) prevailing over the range 0 < x < h.
6.4.5

Multi-Dimensional Extension of OCI Schemes

Various OCI schemes discussed above have been developed in a onedimensional context. In this Section will be considered extensions of the
most promising variants of OCI schemes, namely GOCI and CVOCI to two dimensions. Also issues related to boundary condition implementation and appropriate solution strategies will be examined in detail, as these now assume great
practical significance. GOCI will be discussed first.
i)

Two-Dimensional Extension of GOCI

The particular extension of GOCI adopted in this study is straightforward and is largely based on the work of Lecointe and Piquet.
Considering the two dimensional form of the transport equation (6.56), i.e.,

where the exchange coefficient P0, and the components of velocity, u and v
are presumed constant over the domain of interest for simplicity of
derivation, an OCI scheme can be developed that uses the spatial operators
L(x) and L(y) defined as:

-109

(6.89a)

Ly s

n- i

(6.89b)

Equation (6.88) can then be expressed as:

LX * Ly = S

Using GOCI as developed in Section (6.4.2), the discrete representations

of L and L , given by and , can be related to the discrete
representations of<|>, given by 0, as follows, equation (6.65):

w -x

p,x

J
(6.91a)
or equivalently,

= QX JCX

and

(6.91b)

r 0. ._. ^
(6.92a)
or, equivalently
(6.92b)

The discrete representation of equation (6.90) is then given by:

p x , py
c

(6.93)

or. equivalently

;
S

(6.94a)

where

= Ry 0

(6.94b)

= Ry

(6.94c)

It is important here to note several aspects of equations (6.94).

First, the solution of equations (6.94) requires boundary conditions on 0, as
well as boundary equations for and . Unfortunately, for multidimensional problems it is not always possible to determine the boundary

-110-

equations for and directly from equation (6.88) and the boundary
conditions on 0 as was the case in one-dimensional problems. In these
instances, some appropriately high order representation of Lx and Ly along
the boundaries should be provided.
Secondly, equations (6.94) along with appropriate boundary equations
form a coupled set of equations for 0, JC and . (Compare this to the
one-dimensional case wherejj could be determined explicitly). Therefore,
it would appear that a coupled 3x3 block solution procedure would be required
to solve these equations. One possible alternative is to eliminated and
y

from equation (6.93). This can be accomplished by first multiplying

x -1
both sides of equation (6.94b) by [Q ] and equation (6.94c) by
so that
[Q.Y ] RX >
,
(6.95a)
(6.95b)
Thus, substituting for <jC / and < > from equations (6.95) into equation
(6.94a) yields the following equation for 0.
(6.96a)
or simply,

HW

(6.96b)

Note that [Q] is, in general, a full NxN matrix where there are N
discrete unknown values of 0. This leads to a relatively sparse A matrix
which, unfortunately, has a full band width. Also, the inversions of QX,
y
Q and A are expensive. As a result, the solution of equation (6.94) by
such a procedure requires an excessive amount of computational resources.
Nevertheless, the procedure is useful in that the convectivity of 0. to
its neighbours can be found in the coefficients of the A matrix. This is discussed next.
To determine the connectivity of a node or computational molecule of the
GOCI scheme, it is necessary to consider only the coefficients of matrix A in
the row corresponding to the node in question. Examples.of the connectivity

-111-

of 0,4,4. node in a domain with 7x7 nodes for: Pe = uhx/r0 = vhy/r0 = .

0,1,2,10,100 and are shown in figures (6.8). The most striking feature of
60CI is that the connectivity, in general, spans from boundary to boundary in
both coordinate directions.
In the low Peclet number limit, figure (6.6a),
j
the coefficients alternate in sign, are symmetric and diminish rapidly in
magnitude with distance form 04 j^T. In the high Peclet number limit, the
downstream connectivity is zero, and the upstream connectivity is alternating
in sign but constant in magnitude except at the boundaries. At the
intermediate Peclet numbers, figures (6.8b) and (6.8c), the connectivity is
seen to become progressively asymmetric, weighing more heavily on the upstream
coefficients and rapidly diminishing on the downstream coefficients as Peclet
number increases. Finally, it is noted that for each Peclet number, it can be
shown that the central coefficient of 0.t ,<t. is equal to the sum of the
neighbouring coefficients, although the Scarborough criterion is far from
being satisfied for any Peclet number, i.e., diagonal dominance is violated.
During the course of this study, several iterative methods of 'solution
were developed that treat the coupling between X and jC with various
degrees of implicitness. Generally, solvers which incorporate minimum
implicitness in their formulation performed poorly or even diverged for high
Peclet number problems, even when heavy under-relaxation was introduced. Thus
the following solver, developed specifically for equation (6.94), was found to
display "optimum" characteristics concerning stability and implicitness (hence
computational cost).
If jC is eliminated from equation (6.91a) using equation (6.93), there
results a 2x2 block penta-diagonal equation for 0 and jCy of the form:

rw
0

w
q
0

1-1 .

e"
q
0

y 1+1, j

P X

'

r"

-q"

p,y

where d 1 J - q WS 1 _ l f j

-112-

(6.97)

1.J

l=45<>

+0.009
-0.068

0.415

0.002
*

-0.028

0.275

-1.OOO

0.275

-0.028

0.002

0.000
I r-

t-

-O.OO9

0.153 -1.OOO

0.415 -O.068

+O.OO9

0.153

0.000

6.8a

. Pe = 1.0

6.8b

Pe = 0.0
0.288

-0.133

If

-0.565

0.835

O.OOO

-0.003

0.076

-1.0OO

0.534

-0.133

0.026

o.c

JO

0.000

0.002

-1.OOO

0.835

-O.565

0.

88
I

0.076

O.O02

-O.OO3

0.000

O.OOO

6.8c

Pe = 10.0

6.8d

Pe = 2.0

+0.5

0.462

-1.0

-O.942

00

0.0

O.OOO

-1.0OO

.
0.980

O.OOO

0.0

-O.942

0.462

+1.0

0.0

0.0

1.0

+1.0

-1.0

0
1

0.0

0.0

0.0

6.8e

Pe = 100.0

6.81

Pe

N1683

Figure 6.8

Coefficients Connecting $4^4 with all its Neighbours,

Resulting From GOCI Scheme'Applied to Convected Step Problem
with V at an Angle 0 = 45 to the Horizontal

-113-

Equation (6.97) can be easily solved by a block 2x2 Alternating Direction Line Gauss-Seidel procedure using a block 2x2 TDMA along lines of constant i and constant j. Solution in this manner yields extremely rapid
convergence. In fact for a uniform problem convergence is achieved in one
iteration as Pe - . In the diffusion limit, the number of iterations
required increases but the rate of convergence is usually quite acceptable.
ii)

Two Dimensional Extension of CVOCI

Two-Dimensional extension of CVOCI is readily obtained by integrating

equation (6.90) over the control volume of dimension h by 1 , in the following manner:
h

h . *
x

/~L dydx+y
/ .. Jr\

Jf\

Jf\

L dydx = j
'

^n

*
I Sdydx

(6.98)

^n

Assuming that over the control volume Lx, Ly and S can be approximated by
n X

nV

JL . and S. . the integrals in equation (6.98) can be approximated

i j
'J
' >J
by such that

*
/ L x dydx = f x
J
i j
0

/
I
'O

(6.99a)

(6.99b)

A Sdydx = hS^^

(6.99c)

i~

L y dydx = hn y .

h*
f
J

such that

nV

jcx
1

y
+jC 1

4.

^'.J

equation

(6.93)

Using equation (6.89a), the integral of equation (6.99a) can also be written
as the difference of fluxes between the e and w faces, as in the onedimensional formulation.

-114-

h
/

i
/

h
L'dydx

00

<Pax') +
'ax

v()]
M> /J

*'

dxdy ~. ujgf
*"*"

0" 0
(6.100)

where

ef

Cr

!$

0a

mr

+u
+

dy

dy

**-\

(6.101a)

(6.101b)

0
(cf. equstion 6.13)

Introducing the approximation that a$/ax and * along the e face and
w face are constants, the fluxes are approximated by:
ef * *[r* ff+ u*]e

af +u* 3W

wf ^

Using the one-dimensional analytic solution with an appropriate L profile,

Section (6.4.3), expressions can be obtained for a*/ax and * in terms of 0
andJC, thus enabling J . and J . to be approximated by:

ef *

^"O

wf *

* ""'I

"0 +

*1 +

*qi]e

jC ]

lw

(6.103a)

(6.1035)

where the evaluations or r and q are provided in equstions (6.83), (6.86) snd
(6.87).

Combining equations (6.99a), (6.100b) snd (6.103), the following

relstion between 0 snd X results:

-115-

J1.-J * "l.j
(6.104)
where

./NW

px = (h
..+ /xe /\w.

ij *

V V

In a similar fashion, the integral in equation (6.99b) can be related to

a difference of fluxes through the n face and s face of the control volume.
Then using the solution of (6.89b) to obtain expressions for a*/ay, * and,
y
subsequently, the fluxes in terms of 0 and , equation (6.99b) can be
approximated by an equation of the following form:

u'i,J-i* rlJ *U + rU

(6.105)

where r and q are determined from equations similar to those for r and q in
equation (6.104).
Now equations (6.93), (6.104) and (6.105) represent a set of coupled
algebraic equations which can be solved in the manner described previously for
GOCI. It is important to note, however, that for two-dimensional problems,
due primarily to the use of one-dimensional semi-analytic solutions without
any explicit regard for streamline orientation to the grid, (like SUDS,
Section (6.3)) the solution for * is not necessarily bounded. Nevertheless,
the conservative nature of the discretization may reduce the occurrence and
size of the overshoots and undershoots.

-116-

6.5 IMPROVED SOLVERS FOR A GENERAL SEGREGATED SOLUTION ALGORITHM

Various iterative solvers considered in the following sections are primarily for use with equation (6.11). Equation (6.11) is symmetric, positive
definite with only five diagonals, and has coefficients similar to those that
would arise from an algebraic representation of a typical Poisson equation.
This is in contrast to momentum equations involving both convection and diffusion, i.e., differential transport equations containing both first and second derivative terms, where the coefficient matrix arising from the usual five
point finite differencing is generally neither symmetric nor positive
definite. Discussion will initially consider base solvers with
acceleration techniques to follow.
6.5.1

Implicit Base Solvers for Pressure/Correction Equation of SIMPLE

Derivative Algorithms

Equation (6.11) can be expressed as a matrix equation

Ap = b

(6.106)

where A is the sparse symmetric matrix of coefficients. The direct solution

of equation (6.106) is accomplished via the explicit Choleski Decomposition
method, i.e., the transformation,
A = LLT
where L is the lower triangular matrix and T denotes transpose.
vector is then obtained from:
p = (LT)~1(L"1b)

(6.107)
The solution

(6.108)

However, solving equation (6.107) by simple recursive relations to determine

columns of L introduces entries in L that were zero in the original matrix A,
thus the sparse nature of the equations is lost.
In general, the various iterative means of solving equation (6.106) can
be expressed as:
Me _ = Ap
- b = -r
K
n+1
n
n

-117-

(6.109)

where n is the iteration counter, rn is the residue and en+1,

, = Kpn+l, - pn
the relative error vector. A is a positive definite matrix. The matrix M
depends on the choice of the relaxation procedure. .The basic requirement on M
is that its inverse be easily obtainable and can be related to the matrix A.
The various iterative solution procedures for equation (6.106) utilizing
approximate factorization techniques are derived by constructing M in a manner
such that:
i) its diagonal entries are those of A (point Jacobi method),
ii) the elements of M correspond to the element of A along a line
(line relaxation), etc.
The iterative solution procedures for equation (6.106), as opposed to
the direct solution schemes where M=A, suffer as the inherent implicitness
expressed by equation (6.106) for the variable field is degraded severely.
(26)
However, Stone
in a widely quoted paper devised the iterative Strongly
Implicit Procedure (SIP) which partially overcomes such deficiencies associated with classical iterative methods. In his method, the finite difference
coefficient A is decomposed to comprise the product of an upper and lower
triangular matrix (cf. equation (6.107)). The product of these two arrays is
not exactly the same as the original coefficient array, but means are introduced to partially cancel the effect of spurious non zero entries resulting
from the direct decomposition. A parameter, a, is introduced to control the
degree of partial cancellation. The set of equations generated by the triangular arrays are easily solved and Stone has shown that the method gives
rapid convergence for two-dimensional heat conduction problems involving large
anisotropy.
Incomplete Choleski (1C) decomposition algorithm proceeds essentially in
a similar manner; in fact, it can be shown that when the value of a is set to
zero, i.e. no cancellation, SIP and 1C become algebraically equivalent. The
coefficient matrix is expressed as:
A * M = LOLT -i- E (equivalent to LLT+E)

(6.110)

where D denotes a diagonal matrix and L is forced to have the same sparsity
pattern as A. In equation (6.110) E is a small error matrix. Criteria
associated with the stability of the decomposition is provided by Meijerink
and Van Der Vorst. (85)

-118-

Since M. can easily be evaluated, equation (6. 109) can be written as

M" Apn -

(6

In this form equation (6.111) can be viewed as the solution of:

= Cpn =

(6.112)

with e , as the residue vector. The accelerators described in the follown+1

ing sections can now be applied to the solution of equation (6.112). Base
solvers considered in this study can also be heuristically viewed as providing
the preconditioning necessary for efficient solution as for instance, the conjugate gradient method considered next, would converge in one iteration if
M A=I, where I denotes the identity matrix.
6.5.2 Conjugate Gradient Acceleration (CG)
Conjugate Gradient methods are, in essence, based on variational principles with each iterative step designed to minimize, subject to certain
restraints, a certain non negative quadratic function of the unknowns e.g.,
(the Euclidean error norm e e in equation (6.111)) that vanishes at the
solution. The particular form of the quadratic function to be minimized distinguishes whether a symmetric or a non-symmetric variant of the algorithm is
to be employed. Even though in this study both variants have been considered,
and there exists in the literature recent attempts to develop the algorithm in
a manner appropriate for the general non-symmetric equation, notably that of
/ QC\
Dongarra et al,
the following discussion will emphasize symmetric
equations.
(87)
The quadratic form for equation (6.106) is defined as
Q (p) = -bTp + 1 pTAp

(6.113)

The minimization problem is given by:

; Q (p + at)>=0

(6.114a)

or, equivalently
-bTt +

[tTAp + pTAt]

-119-

(6.114b)

where t denotes a specific search direction for the solution, borrowing

terminology from the closely associated descent methods and a is a numerical
parameter. Invoking the symmetry of A and the fact that

equation (6.114b) is simplified to:

tT[-b -i- Ap] = 0

(6.115)

or Ap=b provided t is not a null vector.

This minimization procedure suggests a solution update of the form
=

P + at

(6.116)

where the search direction t. is said to be A orthogonal if,

t{ A tj - 0

if i j j

(6.117)

In equation (6.117) the mutually conjugate vectors t. are obtained from a

set of arbitrary linearly independent vectors r. by a Gram-Schmidt process.
If,
t1 = r1

(6.118a)

then,
i

"
J-l

,t<

(6.118b)

Utilizing the orthogonal property given by equation (6.117), definition of

Pi+1 is given by:
* 3 t.

(6.119a)

where
At

T
(tj,

-120-

.1>

Atj)

"-H
"

(6.119b)

or

At.)

(6.1190
(r,

.. above denotes the Kronecker delta.

The vectors r. are evaluated using (6.116)
b = r

i ~ Ati

(6.120a)

where, utilizing the orthogonal property leads to:

(tT, r )
a = j

(6.120b)

In the above, expressions like (t , AtJ denote the usual inner

T
1
product of vectors t and At^ . 3 and bi+1 . are scalars, t^ and r.
represent respectively, the current conjugate vector and the residual vector.
Equations (6.116), (6.119) and (6.120) constitute the working equations
of CG methods for symmetric positive definite equations which after N steps
provide the minimum of 0 (p). A specific application of the above algorithm
for use with equation (6.112) without requiring explicit evaluation of the
matrix C is given in the following algorithm form:

+BB.
where a and 0 are given by either

(tl, e^'
(e
a = 1- and B = (t , B.)
(t ,

-121-

or

a =

-.-

and 3 = -

aiiu M

(BJ. B.)

(B\,

and

' V = A V

with suitable starting values for po ,o t and eo.

_
In the above algorithm the preconditioning of A with M in
equation (6.112) has many desirable features. In particular, the eigenvalues
are distributed favorably for the application of the CG method. It is this
preconditioning that is responsible for the significant increase in the rate
of convergence.
For non-symmetric systems including momentum equations, a quadratic form
based on the Eucledian norm can be minimized resulting in a similar algorithm
structure.
6.5.3 Block (Additive) Correction Acceleration (BC)
Block correction and the associated Additive Correction Multigrid discussed in the next section are somewhat different in philosophy and structure
from the previous CG methods and operate on the premise, as shown by
/ tp\
Hutchinson and Raithby,
that the solver convergence is enhanced considerably if the correct global solution level is maintained.
Rewriting equation (6.11) in the following equivalent form:

*'a1Jp1.J-l * b1.J
the equation for the change 4. in level of p along the line 1, figure (6.9),
to ensure that "p" (radially averaged value of p. ) 1s correct relative to
1
.
tJ
its neighbours p
and p. is developed following the procedure of
i 61)
t~\
Settari and Aziz/ ' It 1s assumed that the calculation domain 1s rectangular with m and n nodes 1rvthe x and y directions respectively.
The first step 1s to select blocks containing an Integral number of contiguous control volumes, and to sum the above equation for each control volume

-122-

within the block. Choosing the block to consist of all control volumes in the
i-column results in:

(6.121)

The boundary conditions are assumed to be absorbed into the interior

S
N
P
equations,
so
that
a.
,
=
a.
=
0.
Introducing
0.
.
=
a*
.
LI
c
t
I tn
' J
1 J

- a. _,
ij i
to 4 D.
solution,
p". .+..
1 1 j 1.

- a. . , the left hand side of equation (6.121) reduces

i * J "*" 1 1 ^
.p. .. If p. . is the current best estimate of the
the'improved value of p. after the level shift is p. .=
Substitution of this into equation (6.121) yields:
+ a. 1 _ ] + ^

(6.122a)

where
- P X"*
- E \T
E
i =L1,J: 3i =2^ai,j:
j
j

- W \"^ W
i =2^ai,j
j

.(6.1226)

(6. 122C)

The addition of the equation over the i column block is equivalent to

enforcing the integral equation for p conservation over the column. Solution
of equation (6.122a) for 4. using TDMA, and the subsequent adjustment of
JT. , to obtain p. , results in exact satisfaction of the integral bal'i J .
i J

ance. It is noteworthy that b , the source term which "drives" the

correction, is the sum of the residuals r*. of the p*. . equation over the
i J
J
column. After the level correction, the integral balance is satisfied exactly
(i.e., the sum of the residuals along the line has been forced to zero) and
the task of a point or line iterative solver 1s to ensure local conservation.
A similar equation can be derived for the correction of the level of
p
along rows through additive corrections.
It is worth remarking here that BC formulation as presented above,
assumes no restrictive structure of the parent differential or finite volume
equation and is thus applicable equally to symmetric and non-symmetric form of

-123-

the equations. Equations (6.122), derived for two-dimensional applications,

is referred to as ID Additive Corrections (1DAC) by Hutchinson and
Raithby.
For completeness and ease of reference, Section (6.5.4),
particular extension of the above formulation will be considered next for
three-dimensional applications.
The three-dimensional algebraic equation for the control volume centered
at ijk is of the form:
i
aP

= aE

4.

W
a

4. PN

(6 123

' >

Boundary conditions are again assumed to be absorbed into the Interior

equations. Block correction is implemented by forming blocks of contiguous
control volumes and by summing the equations for the control volumes within
each block. The blocks may be selected in any of a number of ways.
1 DAC : One-dimensional additive correction (10AC) for the level 1 grid in
figure (6.10a) results when blocks are formed from all the control volumes in
the x-z planes (for example). With
_P
E
_W
U
D
i,j,k ~ aij,k " ai-lj,k ai+l,j,k " ai,j,k-l " a1J,k+r

equations like (6.123) are summed over the plane for a j to obtain:

i,
(6.124)

The additive correction is similarly introduced into equation (6.124) in the

manner p.i,J,k
. . K=p".
. .+<$.,
"p. . . is the current best estimate of
i,j,k
j* whereKi,j,k
the solution that 1s to be Improved by adding 6 . A tr1-d1agonal equation
for 6. results for the Level 2 grid in figure (6.10a).
2DAC: A two-dimensional additive correction (2ADC) results when blocks are
formed from all the control volumes in columns that span the calculation

-124-

LEVEL 2

LEVEL 1

Figure 6.9

Grid Layout for ID Block Correction

\ \ i \ \ \ \ \ \
\ \ \ \ x \ \
\

\ \ \ \ \ \ \ \ \

\ \ \ \ \

V \ \ \ \ \
\ \ \ \ \ \

\ \\\ \\\ \\\ \\\ \\\ \\\

V \

v \ \ \ \ \ \ \
\ \ \\
\ \ \ \
\

\l LEVEL

N161

Figure 6.10

Grid Layout for 30 Block Correction. Level 1 is the Fine Grid

-125-

P
U
domain in the z direction. In this case 0, ., k = a^ . .-a. . ._.
~ai i k+1 is introcluced 1nto tne summation of equations like (6.123),
and'the additive correction equation p. . . = p*. . . + 4. . is
lJfK
'JtK
1tJ
substituted. After rearrangement, the equation for . is:
"P

"E

j * *? j*i-l j+ ai j*i j-H + ai j*i j-l+ ^1 j <6-125a)

where
^^ f: < ,
k

*~"

(6.125b)

. V^.W

.E

. \^ N

k
(6.125c)

y
f
a'
i ^
-^P,
^ 1
1, v, ^J
Ifjflv
1J~<>^
k

This equation for 6.

6.10b.

i j

is valid for the level 2 grids in figures (6.10b) and

Equation (6.125a) is a two-dimensional equation that can be solved by

/ to}
iteration. But Hutchinson and Raithby
showed that additive corrections
can significantly enhance convergence for two-dimensional problems. Thus, if
blocks are again chosen to reduce the dimensionality of the additive corrections, solutions are performed on the Level 3A grids to obtain 6. and this
correction is added to 6 ..- This could be followed by a similar solution
1 ,j
on the Level 38 grid to obtain a 4. correction. The use of the three grid
levels in this manner is referred to as 1D2DAC.
6.5.4 Additive Correction Multigrid (ACM)
The block (additive) corrections employed in the previous section have
much in common with the classical multigrid methods of Brandt
and
(88)
others.
Both methods are designed to accelerate convergence by solving
a sequence of problems on increasingly coarser grids. Brandt forms the coarse
grid equations by discretizing the governing equations on each grid and interpolating the fine grid residuals to the coarser grid equations, while the BC
procedure forms the coarse grid equations by asserting integral conservation

-126-

over blocks of control volumes. Brandt interpolates coarse grid corrections

to the fine grid, while BC simply adds the correction in order to preserve the
integral balances. Brandt recommends formation of the coarser grid to include
2x2 nodes of the finer grid (in 2D), while the BC procedure forms the coarse
grid in such a way that the dimensionality of the problem is reduced by at
least one on each higher level (coarser) grid.
The development of the ACM method is now considered in detail, using the
previously established additive correction ideas. It should be noted here
that once the ACM method is adopted, there are no further decisions to be made
regarding the treatment of boundary conditions, the transfer of residuals, or
the interpolation of the dependent variable. The requirement that the sum of
the equations for all fine grid control volumes, that lie within the same
block of a coarser grid, be correct constrains these choices. As previously
discussed, this is equivalent to demanding integral conservation over each
coarse grid control volume.
A Cartesian 7x5 grid is shown in figure (6.11). Rather than form BC
blocks that span the entire domain in one direction, let the blocks be formed
from a number of contiguous control volumes, such as denoted in figure (6.11)
by the heavy lines. To form the BC equation for the k,l block, equations like
(6.121), one for each control volume in the block, are summed. To facilitate
this equation (6.121) is rewritten as:
P
a

*E
P

*W
P

*N

*S

i,J i,J " i,J i+l,J " 1,J i-U

(6.126)

where a. . = a. .-KaT . and /a\ . = 0 if p. , , is in the k,l block and \. . =

i ,j
i , j i, j
i ,j
i"*"i,j
1 ,j
if P. . is not in the k.l block. Summing the N equations for the N control
volumes within the block yields:

(6.127a)
where

*F

*w

*N

*s

u -'ij-Vij- '1*1.1 -i.j-i-'

-127-

Figure 6.11

The Assembly of Control Volumes on the Fine Grid into Blocks

(denoted by heavy lines) that Define the Coarse Grid for ACM

FINE GRID ITERATION

CONVERGED?

LEVEL I <

*
SLOW CONVERGENCE
AND

LEVEL 2 ITERATION
CONVERGED ?

LEVEL 2

N SLOW CONVERGENCE AND

N>2

(FROM LEVEL 3 |
| TO LEVEL (N-1) |

LEVEL V ITERATION

LEVELN <

I
CONVERGED ?

Figure 6.12

Flexible Cycle of Brandt

-128-

if p. . is the best estimate of p. . on the fine grid, the correction

6
obtained on the coarser grid,'is added to pi to obtain an improved
value of p. .. Inserting p . =/p +
into equation (6.127a)
ij
'J
1tj K*

yields:

k i*k i" ak iVi

^ j

^ ^ I

IV j

1+ ak iVi

^ * ^ l y l

^ ^ I

lyl

i+ ak i*k HI* a kVk 1-1* bk

f\ ^ I

K f l ' l

IV^I

^i i

'

'

^ ( I

(6 128a)

(6.128b)

'

(6.128C)

The solution of equation (6.128a) yields the value of

that is added to
each p\ value that lies within the k,l block to obtain'the improved
i, J
estimate, p .
1, J
The blocks may be found in any convenient way. If the grid contains
2 P control volumes in the x and y directions, where P is an integer, it is
convenient to form all blocks from the 2x2 sets of control volumes (20) shown
in figure (6.11). For other grids it is necessary to use unequal block sizes
as also shown in figure (6.11). If, on the other hand, the blocks are formed
from the control volumes along columns or rows that span the domain, the 1DAC
described previously is recovered.
The general BC equation, equation (6.128a), can be solved by iteration,
but convergence can again be enhanced by using BC on a still coarser grid
employing exactly the same method used in the formulation of
equation (6.128), The use of a sequence of coarser grids 1s identical to the
procedure recommended in conventional multigrid methods. Equations on 2x2
grids or smaller are solved by a direct method.
(63^
In the present study the "flexible cycle" algorithm of Brandt
is
adopted. The flowchart is shown in figure (6.12).

-129-

Any iterative base solver can be used with the above acceleration technique, however use of implicit methods including the various approximate factorization techniques considered previously is common.
6.6

ASSESSMENT OF IMPROVED SOLVER CHARACTERISTICS

There are several aspects to be considered when evaluating the performance of the above solvers. These include stability, computational storage re
quirements and computational efficiency. Computational efficiency will be
discussed in detail in Section 7, while the former two Items are considered
next.
i)

Stability

Of the methods considered, only the 1C technique with and without CG

acceleration have been rigorously proven to be stable for the solution of the
symmetric, positive definite pressure/correction equation. For the remaining
methods, SIP with some partial cancellation or any method employing BC or ACI!^
acceleration, no mathematical proof of stability has yet been provided.
Nevertheless, the experience gained from testing and using these methods
indicates that they are in general, stable for the solution of the positive
definite, symmetric set of linear equations for pressure/correction.
ii)

Storage Requirements

The computer storage requirements of each of the methods is somewhat dependent on the manner in which the methods are implemented. For the implementations adopted in the present study, the Storage Unit (SU) requirements for
each of the methods are listed in the table below. A storage unit is defined
as the storage required to store only the dependent variable. The storage
unit requirements listed in the table below indicate the storage required to
store the pressure, as well as the coefficients of the pressure equation.
TABLE 6.1 STORAGE UNIT REQUIREMENTS FOR VARIOUS SOLVERS

Storage Unit Requirements

Base Solver

1C

SIP

10
11
12
20

14
N/A
16
28

Acceleration
CG
BC
ACM

-130-

The various entries in the above table comprise the following SU's:

a) 1 SU for dependent variable

b) 6 SU for pressure/correction equation coefficients
c) 3 SU for 1C
.
.
d) 7 SU for SIP
e) 1 SU for CG
f) 2 SU for BC
g) SU for ACM 2 |(a) + (b) + [(c) or (d)]}
Storage allocations in the above table do not include constants and vectors
(as opposed to arrays).
6.7 CLOSURE

This Section has presented detailed deviations of the previously

selected techniques for further quantitative evaluation, emphasizing Issues of
accuracy, stability and convergence, together with implementat.ional details
including required computational resources. Various relevant attributes of
discretization techniques were discussed in a unified framework along with the
associated questions of boundary conditions, solver suitability for the
particular computational molecules generated by these schemes. Wherever
appropriate, references from the literature were provided to assess a priori
the specific issues of interest. Such information will be readily used in the
next Section to provide a comparative basis to quantitatively evaluate these
schemes in completing selected test cases.

-131-

7.0
7.1

DISCUSSION OF ONE AND TWO-DIMENSIONAL TEST CASES

EVALUATION OF IMPROVED SOLVER PERFORMANCE FOR A GENERAL SCALAR

(PRESSURE) EQUATION

To evaluate the relative computational efficiencies of each of the

solvers presented in Section 6, it would be most desirable to be able to
determine in some formal mathematical fashion estimates of the asymptotic rate
of convergence of each method. Unfortunately, this is, at present, not feasible. Instead, the relative computational efficiencies of the methods can
only be determined through extensive testing and experimentation.
Assessment of individual solver performance in the absence of nonlinearity and variable coupling that generally dominates the behavior of general
flow equations is valuable, as it provides quantitative understanding of the
solver structure. Also, such exercises establish useful guidelines for the
anticipated convergence behavior in the solution of complex coupled equation
sets. Thus, the following sections will critically examine the solver performance in isolation for a variety of linear, scalar (pressure) equations,
while issues related to coupling and nonlinerity are discussed in in
Section (7.2):
7.1.1 Test Problems and Details of Implementation
The following flow problems were adopted to evaluate the solver performance associated with the corresponding pressure/correction solution, Section
(6.1).
i)
Shear driven flow in a square cavity, Re =1000, uniform grid
with 48x48 control volumes, figure (7.1).
ii)
Flow over a rearward facing step with uniform inlet velocity
profile and fully developed outlet conditions, Re.=250, uniform grid with 48 x 48 control volumes, figure (7.2).
iii)
Angled axisymmetric flow through a cannlster, Re =450, uniform
grid with 78x42 control volumes, figure (7.3).
The first test problem was chosen because it 1s generally accepted as a
benchmark, the second because it has a relatively high geometric aspect ratio

-132-

Reh = hV/v= 1000

Figure 7.1

Shear Driven Flow in a Square Cavity

3h

250

0* = 0*

13.6 h

Figure 7.2

Flow Over a Rearward Facing Step

Rer = rV/i/= 450

3r

P = 2JL= 0

3x

15r
N16S3

Figure 7.3

Angled Axisymmetric Flow Through A Cannister

-133-

for the control volume, typical of those often encountered in practice, and
the third problem was chosen because of its geometric similarity to combustion
configurations.
To evaluate the solver performance in isolation, the linear pressure
equations appropriate for the above flow cases were generated and solved in
the following manner:
i)
For each test problem, a converged solution was obtained on a
staggered grid using the upstream weighted discretization scheme
(8)
of Raithby and Torrance
in a finite volume framework.
Relaxation was introduced via the time step multiple, E.
ii)
Using the converged solution of the test problem and E=5.0
(corresponding to usual under-relaxation factor of 0.83), the
coefficients of the equation for pressure estimation of
SIMPLER* 1 * were determined.
iii)
Starting from a zero solution field for each of the methods considered and for each test problem, the solution convergence history (computing effort vs. residual) was monitored and recorded.
In the case of SIP, numerical experiments were performed using values of
a, the partial cancellation parameter, which ranged from 0.0 to 0.9 in increments of 0.1. Results for 1C were obtained using SIP with a=0.0.
For the shear driven square cavity flow, the closed set of algebraic
pressure equations thus generated requires that a pressure be specified at
some point in the field. However, for most iterative methods, the iterative
behavior is enhanced if the pressure is not specified.(15) A prescribed
pressure can then be obtained by adjusting the entire pressure field by an
additive constant. This approach was adopted for all methods tested on the
pressure equation of the cavity problem, except for the methods employing BC
acceleration. Because the BC equations cannot be solved if pressure is unspecified, the pressure was set equal to zero in the upper right corner of
the domain.
7.1.2

Results of Numerical Experiments

The convergence history (1 -norm of pressure equation residuals vs CPU

effort) of each method considered for each of the three test problems is shown
in figures (7.4) to (7.6). In each figure, the norm of the residuals is

-134-

oo
to *

n8

4)
O X

i. CO
O <

>e

V. 4)
O
4-> -O
I/I O I/I
f- 4- O>
Z O. E

O) CL>
O O) O
C -!- >
0> tO
O>
i
i- O) O
O> C t_
>!-->

cue
o mo
<_> u. <_>

L.

a
o>

03 iO) >
I- O)
O 4-

u- re

I- t/> O

a.
en

O
=>
-> 10
l/>
f
i- C O

a> 3 c
ui

u o o
c i o
at u.
o> oo
u c= ^ja> a>
> > X

c <-

O t- 00

w o <

0)

o>
_ o

i i

8
o

-135-

1.0 r

0.1

IC-CG

IIRII
IIRIU

1C

0.01

0.001

\
x
> \
SIP-ACM
IC-ACM
v

SIP-BC
o =0.9

a = 0.5

N1662

20

40

60

80

100

120

CPU EFFORT, t/tsip

Figure 7.6

Convergence History fbr Axisymmetric Flow Through a Cannlster

Problem, 78 x 42 Control Volumes

-136-

normalized by ||R ||, the initial norm of the residuals and CPU effort is
normalized by t , the time required to perform one SIP iteration with no
acceleration. In cases using SIP with partial cancellation, only the convergence histories for the optimal values of a (to within + 0.05) are
presented.
,
Considering first the results for the shear driven square cavity flow
problem, shown in figure (7.4), it is observed that the initial rate of convergence of 1C is particularly good. However, after reducing the residual by
an order of magnitude the rate of convergence of 1C slows dramatically. Combining 1C with CG acceleration significantly improves the asymptotic rate of
convergence but the overall performance of the combination is compromised
somewhat by the non-monotonic reduction in residual. The BC acceleration also
improves the convergence of 1C but not to the same extent as CG. However, the
best acceleration of 1C is provided by the ACM acceleration where the relatively fast initial rate of convergence is maintained throughout.
The results in figure (7.4) also indicate that the Introduction of an
appropriate amount of partial cancellation to 1C, resulting 1n the SIP method,
can also increase the rate of convergence, but only to a limited degree. However when SIP and BC are combined, the resulting rate of convergence exceeds
that of 1C with ACM, but does not exceed the rate of convergence experienced
by the combination of SIP with ACM.
The results for the rearward facing step, shown in figure (7.5), are
only moderately different (qualitatively) from the results for the square cavity problem. The differences lie mainly 1n the relatively poorer behavior of
1C with CG acceleration, which suffers even more from a non-monotonic decrease
in residuals, and SIP with BC acceleration is superior to all other methods
considered.
For the flow through a cannister the results, shown in figure (7.6),
again indicate only minor qualitative differences from the results of the
square cavity problem. The most notable difference which arises is again the
poor performance of 1C with CG acceleration which experiences poor convergence
for a considerable portion of Its history.
7.1.3

Discussion of Results

From the results shown in figures (7.4) to (7.6) 1t is observed that the
initial rate of convergence of all methods 1s good. This favorable behavior

-137-

is likely due to the manner in which 1C and SIP remove the high frequency com(23)
. .
ponent of the error.
However, once this high frequency component of the
error is removed, the convergence, without some form of acceleration, slows
dramatically. This decrease in the rate of convergence is then due to the
poor manner in which lower frequency components of the error are removed. "
Although, it seems that the introduction of an optimal amount of partial cancellation is effective in removing more of the error, the introduction of
acceleration techniques appears to have a more dramatic effect.
The effectiveness of BC acceleration is particularly evident for the
rearward facing step problem, where the solution for pressure has a single,
dominant one-dimensional low frequency mode for which BC is ideally suited.
Similarly, ACM acceleration which is designed to address all frequency components of the solution error, improves the convergence rate of both 1C and SIP
for all test problems. It is also worth noting that for the relatively fine
grids used, the optimal degree of partial cancellation is obtained by using
a=0.9, except with ACM acceleration on the flow through a cannister problem.
In this instance it is postulated that the value of o=0.9, used on all of the
grids, is not appropriate for the coarse grids (for coarse grids, there is
evidence to suggest that the value of a should be decreased).
Finally, the convergence behavior of CG is disappointing. While CG
acceleration has been designed to provide a monotonic decrease in error and to
provide a solution after all the orthogonal base vectors of the solution have
been obtained, there is no guarantee that the intermediate residuals will decrease monotonically and that the rate of convergence at any point will be
fast. In fact/ the present results suggest that the rate of convergence of CG
acceleration is high only after a number of iterations are performed which set
up most of the principal (or dominant) orthogonal base vectors.
Based on the above discussion, the following tentative conclusions
emerge regarding isolated solver performance for single linear equation sets:
i)
Provided the computer storage is available. Section (6.6), ACM
acceleration with SIP employing an optimal amount of relaxation
is generally the most efficient method, but only moderately more
efficient than 1C with ACM.
,
ii)
If it is known, a priori, that the solution for pressure varies
primarily in a single direction, then BC acceleration should be
most effective.

-138-

iii)

If the computer storage requirements of ACM are excessive for

the particular problem, BC acceleration of SIP with an optimal
partial cancellation appears to be an appropriate choice,
iv)
Because the CG acceleration exhibits at times poor convergence
behavior, the method should be used with some caution.
These conclusions are now used in understanding how the choice of the
method used to solve the pressure equation would influence the performance of
a particular segregated method for incompressible flows. Such information
might then also prove useful in comparing the segregated approach to alternative methods, including the coupled approach for incompressible flow problems.
7.2 EVALUATION OF IMPROVED PRESSURE ~ VELOCITY COUPLING ALGORITHMS AND SOLVER
PERFORMANCE IN THE TEACH CODE

The following sections adopt a baseline code to quantitatively study the

implications of employing various convergence enhancement practices for computing isothermal, turbulent, recirculating flows. The code selected was a
variant of the Pratt and Whitney 2-D TEACH Code,
supplied as part of this
contract, incorporating:
i)
Hybrid differencing
ii)
An alternating direction line Gauss-Seidel to solve equations
for pressure, components of velocity and other scalars with a
two-dimensional fives-point operator
iii)
The PISO variant*6* of SIMPLER* 1 * algorithm to deal with the
coupling between pressure and components of velocity
iv)
k-c turbulence model.
Whereas .the previous sections examined the solver performance in isolation for single linear pressure equations, the relevant details of the overall
segregated solution algorithm for incompressible flows is considered
next.
7.2.1

Details of Implementation

The convergence and efficiency of a typical segregated incompressible

solver is primarily dictated by, Sections (3.4.3) and (4.3):
i)
The nature of the approximations introduced to deal with the
crucial pressure ~ velocity coupling,
ii)
Effective solution of the resulting pressure/correction equation.

-139-

These issues will be discussed in detail with reference to the particular means of incorporating the associated convergence enhancement algorithms
in the TEACH code.
To improve the coupling between pressure and velocity components, the
SIMPLEC
algorithm was, employed. The SIMPLEC algorithm is a more consistent implementation of the original SIMPLE method resulting in considerably
improved convergence behavior. Along with the implementation of SIMPLEC, the
baseline code was modified so that mass conserving velocity components are
always used when discretizing all transport equations. In the baseline code,
coefficients of the discretized v momentum equation were based on u* velocities, which, in general, do not satisfy mass conservation, Section (3.4.3).
The standard implementation of PISO and SIMPLEC, as reported in the
literature, are such that under-relaxation of momentum equations controlled by
a, the under-relaxation parameter, is required. The under-relaxation, in
turn, results in a relatively large number of coefficient updates (iterations). To reduce the number of coefficient updates, the SIMPLEC cycle of uv;
v* and pressure calculation is applied repeatedly using the same sets of
coefficients of the momentum equations. The required under-relaxation of
SIMPLEC is accomplished by adding
ar0/aolnrLtL
-u ) to the u momentum
CTUmcr (u
1
a a
v
v
ne v
equation and p/ sTup|rr ^
~ ) * *
momentum equation where aS
and a^, are the central coefficents of the corresponding momentum
equations, o_olnrLtL
T U _.__ is the SIMPLEC under-relaxation parameter and the
superscript n is used to denote the current SIMPLEC iteration. Termination of
the SIMPLEC cycle is based on the reduction of the u* and v* mass conservation
residuals and is controlled by the value of the residual reduction factor,
a
D The result is that, after repeated applications of SIMPLEC,
SIMrLtL
pressures and velocities better satisfy mass and momentum conservation and
that the under-relaxation of the momentum equations controlled by a, can be
substantially reduced, thereby reducing the number of coefficient updates
required. This reduction in coefficient updates is particularly important for
the more expensive discretization schemes like MW-SUOS and LP-SUDS, where the
cost associated with the updating of coefficient is relatively high.
Regarding the convergence enhancements associated with the solution of
the pressure/correction equation, SIP with CG, BC or ACM acceleration techniques were implemented in the baseline code. Each of the options was tested,
however, guided by the conclusions of the previous section only SIP with BC or
ACM accelerations were used in the demonstration problems.

-140-

To further enhance convergence, the residual reduction convergence

criterion, controlled by the residual reduction parameter, Y i
Section (4.3), was implemented in the code to establish an efficient termination procedure for the iterative solution of the pressure correction equation. In addition, the approximate factorization of the pressure correction
equation is stored, i.e., approximate factorization is performed only when the
coefficients of the pressure correction equation are updated. This is particularly useful in PISO where solutions of similar pressure correction equations
with differing driving source terms are required in the two stages. The only
exception to this artifice arises with SIP-CG where the initialization of CG
requires that the decomposition be performed in both stages of PISO.
Finally, as discussed in Section (6.5), SIP with BC or ACM acceleration
is equally applicable to symmetric and non-symmetric equations. Therefore, it
is possible to use these methods for the solution of momentum equations as
well as scalar transport equations. However, as these equations are generally
dominated by convection and/or source terms, algebraic solutions are readily
obtained without using any acceleration techniques. Therefore, SIP with BC or
ACM was not implemented for these equations.
7.2.2

Test Problems and Procedure

To study the impact, the above practices are likely to have on partial
computations of flow problems with associated nonlinearity and variable coupling, solutions were obtained to the following demonstration problems:
i)
Shear driven flow in a square cavity, Re =1000 using a uniform
grid with 10x10 control volumes, figure (7.1).
ii)
Flow over a rearward facing step, Re =250 using the boundary
conditions as in Section (7.1.1) and a uniform grid with 10x6
control volumes, figure (7.2).
iii)
Coannular, nonswirling, turbulent flow in a sudden expansion
geometry. The inlet conditions are presented in Table (7.1). A
non-uniform grid with 21x20 control volumes, similar to that
described in (5) was used.

-141-

TABLE 7.1

INLET CONDITIONS FOR THE COANNULAR TURBULENT FLOW PROBLEM

Variable

Inner Flow

Outer Flow

u (m/s)

0.596

l.>74

v (m/s)

0.0

0.0

k (m2/ S 2)

1.776x10-3

1.514x10-2

e (m2/ s 3)

1.52x10-2

3.78x10-1

Computations were performed using the familiar Hybrid differencing to

obtain the dicretized algebraic equations. Use of such coarse grids for the
above demonstration problems 1s intentional in that, on finer grids the
enhancement of convergence and accuracy provided by the above practices is
generally more dramatic. There are several reasons for this Including:
1)
On coarse grids, the increased computational effort required to
set up SIP with acceleration may not be offset by the reduction
in the number of iterations required to solve the pressure correction equation,
ii)
On coarse grids, the performance of SIMPLER/PISO can be better
than the performance of SIMPLEC.
iii)
The cost of coefficient updates of the Hybrid discretization
scheme is quite low. As a result, the effectiveness of repeated
applications of SIMPLEC for fixed momentum coefficients will be
diminished.
The various solution parameters adopted for the test problems are summarized in Table (7.2). In all the test problems studied solution was assumed
converged when the corresponding normalized maximum residual in the field was
reduced to 0.005. The CPU requirements quoted in Tables (7.3) through (7.6)
refer to a MASSCOMP Series 500 minicomputer.

-142-

TABLE 7.2

7.2.3

PARAMETER SETTINGS FOR CONVERGENCE ENHANCEMENT TESTS

SIP Partial Cancellation Parameter

0.5

Pressure Correction Residual Reduction Factor, yp

0.25

SIMPLEC Under-Relaxation of Momentum, <*siMPLEC

0.7

SIMPLEC Residual Reduction Factor, V$IMPLEC

0.4

Results of Numerical Experiments

For the shear driven square cavity problem, the number of coefficient
updates and CPU requirements for the baseline code, the baseline code with
SIP-BC and the baseline code with repeated SIMPLEC and SIP-BC are presented in
Table (7.3).
TABLE 7.3 CONVERGENCE ENHANCEMENT TEST RESULTS FOR SHEAR DRIVEN
CAVITY PROBLEM, GRID = 10 X 10

Solver

Under-Relaxation
Parameter, a

Coefficient
Updates

Total CPU
Requirements
(Seconds)

Baseline

0.5

86

305

PISO/SIP-BC

0.5

86

182

SIMPLEC/SIP-BC

0.92

20

101

The above results indicate a 40 percent reduction in overall CPU requirements with the introduction of SIP-BC. The introduction of repeated
SIMPLEC allows a much higher value of the under-relaxation factor a and
results in a 77 percent reduction in the number of coefficient updates, as
well as an additional 44 percent reduction in CPU requirements.

-143-

For the rearward facing step problem, the number of coefficient updates
and CPU requirements for the baseline code, the baseline code with SIP-BC,
repeated SIMPLEC, and the baseline code with SIP-ACM and repeated SIMPLEC are
given in Table (7.4).
TABLE 7.4 CONVERGENCE ENHANCEMENT TEST RESULTS FOR REARWARD
FACING STEP PROBLEM, GRID = 10x6

Solver

Under-Relaxation
Parameter, a

Coefficient
Updates

Baseline

0.5

26

37

PISO/SIP-BC

0.5

25

46

SIMPLEC

0.8

13

36

SIMPLEC/SIP-BC

0.8

12

34

SIMPLEC/SIP-ACM

0.8

13

38

Total CPU
Requirements
(Seconds)

As will be seen from the above table, for the coarse grid used for this
problem, there is little or no improvement in CPU requirements with the introduction of SIP with acceleration or repeated SIMPLEC. In fact, the introduction of SIP with acceleration is a disadvantage for this coarse grid.
However, it is encouraging to note the 50 percent reduction in coefficient
updates that result from the introduction of repeated SIMPLEC and a higher
value of a. Also note some reduction in CPU requirements with SIP-BC used in
conjunction with repeated SIMPLEC.
For the turbulent coannular flow problem, the number of coefficient
updates and CPU requirements for the baseline code, the baseline code with
SIP-ACM the baseline code with repeated SIMPLEC and the baseline code with
SIP-ACM and repeated SIMPLEC are presented in Table (7.5).

-144-

TABLE 7.5 CONVERGENCE ENHANCEMENT TEST RESULTS FOR TURBULENT COANNULAR

FLOW PROBLEM, GRID = 21x20, TRUE VALUE OF vm)n * 0.117(5)

Parameter, a

Coefficient
Updates

Baseline

0.5

178

40

PISO/SIP-BC

0.5

Unstable

SIMPLEC

0.9

110

37

SIMPLEC/SIP-ACM

0.9

109

34

Solver

. , Under-Relaxation

Total CPU
Requirements
(Seconds)

Considering, first, the impact of using SIP-ACM for the solution of the
pressure correction equation, it 1s observed that PISO becomes unstable with
the improved method for solving the pressure correction equation. Similar
results were obtained using SIP-BC. The reason for this behavior is that for
the turbulent coannular flow problem PISO becomes unstable on a 21x20 grid
using o=0.5. In the baseline code, use of the recommended 5 iterations of the
Alternating Direction Line Gauss-Seidel procedure for the pressure correction
equation, which is considerably slower than SIP-ACM, introduces sufficient
relaxation to stabilize PISO. It can also be seen from the above table that
the Use of repeated SIMPLEC results 1n a 40 percent reduction in the number of
coefficient updates. However, with the increased costs of repeated SIMPLEC,
total CPU requirements were reduced by only 9 percent. Repeated SIMPLEC
together with SIP-ACM results in a 15 percent reduction in CPU requirements.
As indicated previously, dramatic reductions in CPU requirements resulting from the use of repeated SIMPLEC are expected, when CPU intensive schemes,
like MW-SUDS or LP-SUOS, are employed for discretization. To illustrate this
issue, numerical solutions to the turbulent coannular flow problem employing
MW-SUDS on the 21x20 grid were obtained using the baseline code, the baseline
code with SIP-ACM, the baseline code with repeated SIMPLEC and repeated
SIMPLEC with SIP-ACM. The number of coefficient updates and CPU requirements
for each solution are presented 1n Table (7.6).

-145-

TABLE 7.6 CONVERGENCE ENHANCEMENT TEST RESULTS FOR TURBULENT

COANNULAR FLOW PROBLEM USING MW-SUDS, GRID = 21x20 TRUE VALUE OF Lr * 6.1

Under-Relaxatlon
Parameter, a

Coefficient
Updates

Baseline

0.5

272

160

PISO/SIP-BC

0.5

306

174

SIMPLEC

0.9

160

119

SIMPLEC/SIP-ACM

0.9

137

92

Solver

Total CPU
Requirements
(Seconds)

The results Indicate that the introduction of repeated SIMPLEC

to the
i
baseline code yields a 41 percent reduction in the number of coefficient
updates and a corresponding 26 percent reduction in total CPU requirements.
Combing SIP-ACM with repeated SIMPLEC results 1n an additional 14 percent
reduction in the number of coefficient updates and an additional 23 percent
reduction in total CPU requirements. Compared to the baseline code, SIMPLEC .
with SIP-ACM results in a 43 percent reduction in total CPU requirements.
Finally, it is noted that the introduction of SIP-ACM 1n PISO results in an
increase in both the number of coefficient updates and the total CPU
requirements. This behavior again highlights the unstable nature of PISO.
7.2.4

Summary and Conclusions

Convergence enhancement and efficiency for the nonlinear, coupled transport equations of incompressible flow were examined above using an improved
pressure ~ velocity coupling algorithm together with a variety of appropriate
solvers for the pressure correction equation. Although more extensive testing
and use would be required to arrive at any general conclusions, the results
for the test problems considered clearly delineate the complex relationship
between the structure of a segregated solution algorithm and the efficient
solution of the pressure field to yield zero divergence for mass. The suitability of a pressure solver for use with a specific incompressible flow
algorithm cannot be overemphasized, as PISO was shown to be unstable for some

-146-

of the test problems examined. The above study also highlighted the efficiency aspects of an improved solution algorithm when used with accurate but
more CPU intensive discretization schemes. However, the problem is generally
much more involved including issues of boundedness, stability and convergence
characteristics.
Based on the results of the above test problems, the following conclusions are stated:
i)
The introduction of an accelerated SIP to solve the pressure
correction equation can significantly reduce CPU requirements;
up to 40 percent reduction for laminar flows and 20 percent
reduction for turbulent flows.
ii)
The introduction of repeated SIMPLEC significantly enhances convergence, reducing both the number of coefficient updates and
CPU requirements. Because of the relatively high cost of coefficient updates for CPU intensive discretization schemes, like
MW-SUOS and LP-SUDS, the introduction of repeated SIMPLEC has,
generally, the largest impact on the CPU requirements of such
schemes.
7.3

EVALUATION OF IMPROVED DISCRETIZATION SCHEMES

The governing transport equations for fluid flow in arbitrary configurations of engineering interest express a delicate balance between the various
influences of convection, diffusion and source terms. Section (6.1). Different zones in the flow field emphasize the influence of one or more of these
transport mechanisms at the expense of others.
Due to the extreme nonlinearities involved in the above equations,
especially at high Reynolds numbers, very few exact (analytical) solutions can
be found in the literature. These usually relate to simple, idealized (very
often linear) problems, constructed in a manner to examine constituent effects
in more comprehensive equations like the Navier Stokes. The boundary and initial conditions associated with the above solutions are generally specified as
analytical functions. For purposes of developing numerical solution techniques, these exact solutions provide an invaluable tool for evaluation purposes, however, idealized the physical model problem might be.
Very often in the literature, numerical solutions embodying various constituent physical models are compared with fairly complex experimental configurations. Description of the initial and boundary conditions is far from

-147-

adequate for numerical model verification. The significant influence boundary

conditions exert on the solution behavior has been amply stressed by several
(2-4)
researchers.
Subsequently, the rather disappointing picture that
emerges a result of comparing numerical solutions with experiments is further
aggravated, as it now becomes an extremely difficult task to identify and "
isolate the various sources of discrepancy. Among these are the numerical
solution techniques, physical modelling assumptions.
Thus, for purposes of evaluating the previously discussed discretization
techniques, this study has adopted the strategy that assessment exercises are
confined predominantly to examining the performance of discretization schemes
in the light of exact solutions. These also include thoroughly exercised
(documented) model numerical solutions. Where appropriate, the performance of
a scheme was further studied in more demanding flow problems including nonlinearity and variable coupling. A typical scheme's performance is also
examined with respect to the particulars of the alternative techniques considered. The assessment criteria, presented in Section (4.4), for such comparison exercises were manipulated to comprise the following attributes with a
view towards three-dimensional applications:
i)
Best accuracy on coarse grids. Even with advanced computers,
many three-dimensional calculations will still be limited to
grids which are, at best coarse.
ii)
Robustness and stability. For many three-dimensional computations it is important to ensure that numerical solutions can be
obtained. Particular reference is made here to k and c
equations,
iii)
Efficiency. Improvements in accuracy must be attained without
excessive computational requirements.
The performance of the schemes will now be discussed in detail. Unless
otherwise stated, due to the extended computational molecule, a nine point
Alternating Direction Line Gauss-Seidel procedure was used effectively, even
in the presence of mild negative coefficients introduced by SOU or LP-SUDS, to
solve the linear algebraic equations.
7.3.1

Second Order Upwind (SOU) Differencing Scheme

To evaluate the SOU discretization scheme, SOU 1s applied to four scalar

transport model problems, as well as several flow problems. The scalar transport problems are all described by equation (6.1) in Cartesian form, i.e.,

-148-

fj (pui) + fy (pvi) = Jj (Ii|J) = ^ (raf*) + S^x.y) Equation (6.1)

and include the following specific cases:

i)
The uniform, angled flow, convected step problem chosen to
examine the behaviour of the scheme with the flow at various
angles to the grid, figure (7.7).
ii)
The uniform angled flow past a unit source chosen to evaluate
the performance of SOU in the presence of a source, figure (7.8).
iii)
The problem of curved flow through a rectangular domain (the
so-called Smith-Hutton
problem), adopted to examine the
behavior of SOU in the presence of flow curvature, figure (7.9).
iv)' Problem (ii) with a distributed source give by:
S(x,y) = Pu(&x-6x2)(3y2-2y3) + Pv(3x2-2x3)(6y-6y2)
-

(? 1}

2
rj(6-12x)(3y
-2y3) + (3x2-2x3)(6-12y)]
0

with p=l, u=2, v=l..

The solution for 0 in this case is given by:
0 = (3x2-2y3)(3y2-2y3)

(7.2)

For each problem, SOU was applied only to cases where the flow was at an angle
to the grid where the performance of SOU is expected to be the worst.
For the problem of scalar transport of a step, three cases were considered: Case I, Pe =250, 8=45 on a uniform 5x5 grid; Case II, Pe =250,
6=30.9 on a uniform 5x5 grid; and Case III, Pe =250, e=30.9 on a uniform
25x25 grid.
For Case I, the results shown in figure (7.10a) indicate that for 8=45
the SOU results exhibit an excessive smearing of the step profile as compared
to the 'exact' solution obtained by Huget.
In fact, comparing the results to previously obtained.solutions, the smearing of SOU is comparable to

-149-

1
<f>- 0

Figure 717

Convected Step Profile Problem

'77-s = 10

/L/L-19

-<.Q

4>-o

Figure 7.8

Uniform Flow Past a Single Volume Source

- 1-tanh (10)

> - t-tanh (10)

u 2y (1 - )
v 2X(1 - )

P..

INLET, <- 1 + lanh [lO(2 x +1)]

Figure 7.9

OUTLET, A?. .

Smith - Hutton Problem

-150-

1.00
0.75

a)

e.50
0.25

0.00
0.0

0.5

1.0

r/L

1 .00

b)

0.75

0.50

EXflCT

NU-SUDS
---- sou

0.25

0.00
0.

0.5

1.0

T/L

1.00

0.75

0.50

c)

nu-suos
0.25

L P - S U D S , EXACT
SOU

0.00

0.0

0.5

1.0

T/L

N1689

Figure 7.10

Centre-line Profiles of 0 for Convected Step Problem

-151-

the smearing of the Upwind Differing Scheme (UOS). Also shown in figure
(7.10a) are the results of Mass Weighted Skewed Upstream Differencing
(MW-SUDS), which exhibit considerably less smearing of the step profile. Case
II provides a comparison of SOU and MW-SUDS for 9=30.9 where MW-SUOS exhibits
more smearing than for e=45. The results shown in Figure (7.10b) Indicate
that for the coarse 5x5 grid SOU smears the step profile more than MW-SUDS.
To determine the behavior of SOU on moderate grids, a grid of 25x25 was used
in Case III. The results shown in figure (7.10c) indicate that SOU does not
smear the step profile as much as MW-SUDS. However, the SOU result exhibits
one percent overshoot, while the results of the Linear Profile Skewed Upstream
Scheme (LP-SUDS) exhibited no overshoots. Section (7.3).
For the problem of scalar transport with a unit source, the only case
considered was that of 0=45. The results shown 1n figure (7.11) indicate
that SOU smears or spreads on the influence of the unit source, as compared to
(91)
the "exact" solution obtained by Stubley.
For the problem of scalar transport of a prescribed Inlet profile in a
prescribed flow with curvature the case Pe=10 was considered. The results
shown in figure (7.12) indicate that, at the outlet, SOU exhibits smearing of
the profile comparable to that of MW-SUDS as well as overshoots and undershoots as large as LP-SUDS.
To evaluate the rate of convergence of SOU with grid refinement, four
uniform grids 4x4, 8x8, 16x16 and 32x32 were used to obtain SOU solutions for
the scalar transport problem with a distributed source term. The resulting
RMS error and the rate of convergence are given in Table (7.7). The results
indicate that the rate of convergence of SOU approaches 2 only as the mesh
becomes very fine.
TABLE 7.7
Grid (N)

RATE OF CONVERGENCE OF SOU

RMS Error

Rate

4x4

3.194x10'*

8x8

1. 2216xlO- z

1.39

16x16

3.665xlO- 3

1.74

32x32

9. 859x1 0~ 4

1.89

-152-

1 .00

0 = 4S

0.75

<J)

y/L = 0.5

EXflCT

Grid: 19*19

SOU

0.50
0.25

0.00
0.0
Figure 7.11

0.5

1.0

X/L

Centre-Hne Profile of 0 for Unit Source Problem

2.0

EXflCT
NU-SUDS
LP-SUDS

1.5

SOU
1 .0

0.5

0.0

0.0

0.5

1.0

N1688

Figure 7.12

Outlet Profiles of 0 in Flow with Curvature

-153-

In the above table, the convergence rate 1s defined as:

i

FRMS Error!
coarse
[RHSNError]fJne
KSiuG

(7.3)

.u

coarse
In addition to the above scalar transport cases three laminar flow test
cases, namely the shear driven flow in a square cavity, flow over a rearward
facing step and angled flow into a suddenly expanding pipe were executed. The
incompressible flow algorithm used was the PISO variant of SIMPLER algorithm
and an Alternating Direction Line Gauss-Seidel scheme was used to solve equations for pressure (19) and components of velocity.
For the shear driven cavity problems, figure (7.1), four global Reynolds
numbers, Re. , of 100, 400, 600''"and 1000 were investigated using a uniform
grid of 40x40 to examine the changing flow structure in the cavity due to
increasing influences of convection. In the case of Re =100 both SOU and
hybrid differencing are in agreement with the "exact" numerical solution of
(92)
Burgraff
for the axial velocity profile at the vertical centre-plane, as
displayed in figure (7.13). However, as the Reynolds number increases and
convective influences begin to dominate, the additional accuracy afforded by a
higher order scheme like SOU, as compared to first order schemes, becomes
apparent. Nevertheless, there is still a considerable smearing of the centreline axial velocity profile as is apparent from figure (7.14) for Ren=600.
None of the above solutions suffered from the severe solution difficulties
reported by Vanka (93) for high Reynolds number cases, but displayed minor
overshoots and undershoots (up to 4 percent) for Ren=1000, Section (6.2.1).
Application of SOU to predict the flow details over a rearward facing
step, figure (7.2), yielded the global recirculation zone length to be 5.67
step height employing a non-uniform grid distribution of 62x38 and a tophat
(constant) inlet velocity profile. The Reynolds number based on step height
was 250, for which the experimentally measured reattachment length is given by
(94)
Durst
as 6.3 step heights. Table (7.8) compares the reattachment length
LK =xK /h, for various alternative differencing schemes and grids.

-154-

3
e
O 3

MO -M
<U
C
1- O

;- no o

<*- O)
P f O

I +J o
0.r- ifl

1- o

*-> O II

CO

Ol O) 0)

o> oe

C
r- O

u- in
03^-:
o
10 r ||

O> Ub^
r>j' 0)
r- -O O. E
VI- 0)
3

o>

t- cn c o
Q- c o >
< f-

"O
t/>
0) L. C O
+J O 10 4(O U- O.-M
IX C

o> >uj o

O E
U_ r-

3 C>
0>

|1
0)

- 4- O

^ <o cj
O O

I- (^

O.

X
1- O

A) O ^"

I *-> O
0> v- r
i. O
- O II
C i
0> 0) 0>
O > Qt

CO

-155-

C +->
U
O) -i- C
C9 U 0> <

o -o ^r

~\ ^- ^ j<
O

0> 3 CO

TABLE 7.8 COMPUTED REATTACHMENT LENGTH LR FOR VARIOUS ALTERNATIVE

DIFFERENCING SCHEME FOR THE REARWARD FACING STEP (Ren=250)

Grid

Hybrid

QUDS(69)

Tophat 1

10x6

2.5

5.32

2.2

Tophat 2

20x12

3.14

5.25

3.95

Tophat 3

40x24

4.52

5.35

5.78

Tophat 4

62x38

Tophat 5

78x48

Inlet Profile

BSUOS2(5)

SOU

5.67
5.53

5.75

5.98

- -

It will be seen from above table that the performance of SOU indicates
a
/ eg \
behaviour very similar to that of QUICK differencing of Leonard
(better
bounded solutions, however) and far superior as compared to Hybrid differencing.
Concerning the problem of angled inflow into a sudden expansion type
geometry, the predicted flow field by SOU (non-uniform grid 78x44, Re. =450,
inlet angle=30) displays qualitatively the expected correct behavior, i.e.,
the multiplicity of recirculation zones and the complex interaction between
the various zones, figure (7.15). In this case, however, there is neither any
experimental data nor carefully performed "exact" numerical predictions
available for comparison.
7.3.2

Variants of Skewed Upwind Differencing Schems (SUDS)

To demonstrate the applicability of the Linear Profile and Mass Weighted

SUDS (LP-SUDS, MW-SUDS) and examine their performance, a number of twodimensional scalar transport and flow problems have been considered. These
will be considered in detail now.
Scalar Transport Problems
Three scalar transport problems, similar to SOU, were chosen to evaluate
the variants of SUDS schemes incorporating an implicit determination of the
interface values of variables using a simple LU decomposition. These are:
i)
The uniform, angled flow convected step problem
11)
The uniform angled flow past a unit source
iii)
The problem of curved flow through a rectangular domain, the
Smith-Hutton problem.

-156-

For problem (i) the relevant problem parameters examined Include

Pe,=250, four flow angles, e=0, e=11.3 (tan'^O^)), 0=30.9 (tan'1
(0.6)) and 9=45 and a variety of uniform grids including 5x5, 15x15 and 35x35
control volumes.
Figure (7.16) displays the results at the domain centreline, x/L=0.5
obtained using LP-SUDS and MW-SUDS and compared to the "exact" results
obtained by Huget
for problem (i). Also shown on the figure are the
coarse grid Exponential Differencing Scheme (EDS) and SUDS results obtained by
Huget.'30'
In general, the results indicate that, as the angle of the flow to the
grid is increased, MW-SUDS exhibits some false diffusion, but to a much lesser
degree than EDS and LP-SUDS exhibits some spurious spatial oscillations, but
usually much less than those observed for SUDS.
For problem (ii) three flow angles were considered, 8=0, 22.5 and
45. Centre-line (y/L=0.5) profiles obtained using 19x19 uniform control volumes are shown in figure (7.17). Also shown are the exact results obtained by
Stubley. (91) In general, the results indicate that:
a) Both LP-SUDS and MW-SUDS benefit from source term influence. In
particular, accounting for the influence of the source term reduces
the overestimation of maximum 0 and shifts the location of maximum 0
downstream and closer to the true location.
b) Due to the false diffusion introduced by MW-SUDS the overestimation
of maximum 0 by MW-SUDS 1s less than that of LP-SUDS. However, at
the downstream boundary, the false diffusion of MW-SUDS results in
overestimation of 0.
For problem (iii) illustrated in figure (7.9), the flow transports a
slightly diffused step profile of 0 into the computation domain through the
left side of the bottom boundary and it leaves the domain through the right
side of the bottom boundary with the 0 profile determined by the Peclet number, Pe. Using a 20x10 uniform mesh, numerical solutions are obtained for
6
five Peclet numbers, Pe=10, 100, 500, 1000 and 10 . The profiles along the
outlet portion of the bottom 0 < x < 1 are shown in figures (7.18). In general, the results for this problem with flow curvature Indicate that:
a) for low values of Pe, LP-SUDS and MW-SUDS results are in close
agreement.

-157-

_ q

o>
X)

o
a.
to
o

Q)

O
0)

o
o
o

<4-

a
*o

</i

4)

0)

I on

2! I
+> 0)

c .c

0) <-)
O CO

0)

a>

-158-

oo
-. p

oo
o>

o
l_

Q.

Q.
O)

oo
o
0>

<D
>

c
o

I/I

o>

o
Q.
0)

c
O> O)

<- e
*->

0>

C -C
0) U

o oo

o
u

o>
O>

-159-

i/t
0)

0)

fi
o
00

t/1

0>
X

0)

o
*

3
O

m
rd

in
6

in
N
6
(.
O

^.
1/1
0)

a.
a>
i
a>

+->
c
a>
u

a>
3
O>

in
!;

in
6

in
CM

-160-

0>
JZ

10
4)
O

a.

1
p
CM

in

T-

i
m
6

P
0

e-

VI

a>

</)

D 0

o
M

o
a.

7
7
'

*J
at
3
^
O

-V

CO

a>

in
O

p
CM

-161-

in
o

0)
.O
O

o.

I/I

O
Q.

(^

+J O)

a) e
r- 0)
(-> f
3 U
O 01

c
O
u
oo

3
M

in

p
^

in
d

-162-

o
d

o
o

b) as Pe is increased, the false diffusive nature of MW-SUOS becomes

6
evident. It is also worthy to note that for Pe = 10 and 1000,
MW-SUDS results do not differ significantly from the results for
Pe = 500.
,
c) as Pe is increased, LP-SUDS exhibits no false diffusion but over3
'
'
'
' '
shoots become apparent for Pe > 10 with a maximum overshoot of '..
6.25 percent occuring for Pe = 10 .
All the above scalar transport problems employed an implicit formulation
for the determination of interface variable values. To illustrate the applicability of explicit linear profile and mass weighted SUDS, Section (6.3.4),
numerical solutions to the same problems were used to demonstrate the performance of explicit skew schemes. For the convected step, unit source and
Smith-Hutton problems, the explicit skew results are almost identical to the
implicit skew results. However, the explicit skew results required less than
50 percent of the computational effort required by the implicit skew schemes.
Using the current implementation, coefficient assembly costs for LP and
MW-SUDS are estimated to be four times the assembly cost for the baseline
Hybrid differencing.
For the problem of scalar transport with distributed sources in a uniform velocity flow (not discussed above), the explicit skew results are again
similar to the implicit skew results. The exception, however, is the performance of the LP-SUDS with source term influence. As shown in Table (7.9), the
rate of convergence of LP-SUDS-STI, where suffix STI denotes source term
influence, is now 2. This second order rate of convergence is not a result of
the explicit skew approach. Instead, it is a result of using a first order
approximation to the source term influence in the vicinity of boundaries. In
the implicit skew implementations, a zero-order approximation to the source
term was used in the vicinity of the boundaries.
TABLE 7.9

Grid (N)

RATE OF CONVERGENCE OF LP-SUDS-STI

RMS Error

Rate

2.022 x 10*

5.0779 x TO'3

1,99

16

1.2657 x 10~3

2.00

32

3.0889 x 10~4

2.03

-163-

In summary, explicit LP-SUDS and MW-SUDS require less computational

effort and yield numerical results that are almost identical to the corre*'
'' \
'
spending implicit schemes. Also, by ensuring that a first order approximation
to the source term influence is used throughout the domain including near
boundaries, LP-SUDS-STI has been shown to be a scheme with a second order rate
of convergence.
1

Fluid Flow Problems

The previous laminar flow demonstration problems, i.e., shear driven .
flow in a square cavity and flow over a rearward facing step, as well .as a
turbulent coannular flow were chosen to flow evaluate the variants of SUDS for
flow problems. The emphasis here is on the discretization of the momentum,
conservation equations with and without accounting for the pressure gradient
source term influence. The solutions discussed below were all obtained using
repeated SIMPLEC with SIP-BC or SIP-ACM and an implicit/explicit formulation
for the interface value of the relevant variable.
.;.
For the driven cavity flow, Re. =1000, two grids were used to obtain
numerical solutions, a uniform 10x10 mesh and a uniform 20x20 mesh. From
these numerical results, obtained using an implicit interface value formulation and SIP with BC, values of minimum normalized stream function,
y...

mm
were determined and tabulated in Table (7.10). Also tabulated are the EDS and
SUDS results obtained by Huget.(30)
TABLE 7.10 VALUES OF <|rm1n FOR DRIVEN CAVITY PROBLEM, EXACT VALUE
OF
~ -0.117(95)

~^\^^ Grid
10x10

20x20

EDS
SUDS

-0.0494
-0.0370

-0.0677
-0.0577

MW-SUDS
LP-SUDS

-0.0457
-0.0401

-0.0643
-0.0599

MW-SUDS-STI
LP-SUDS-STI

-0.0610
-0.0551

-0.0786
-0.733

Scheme

^^-~^

-164-

From the results tabulated, it is seen that even the best solution obtained by using MW-SUDS-STI differs considerably from the computations of Ghia
et. al.
In noting similar discrepancies, Huget
suggests that they
are due to the poor treatment of shear in the vicinity of the moving lid.
There are at least two ways to overcome this difficulty; one is to introduce
(30)
higher order representations for boundary conditions,
and another 1s to
introduce influence points to evaluate the mass fluxes in the representation
of mass conservation. This latter approach has been demonstrated by Raw
to result in more accurate results for the driven cavity problem. In this and
other flow problems in this section, the source term influence of pressure in
momentum conservation, when it was accounted for, was implemented using a
deferred correction approach.
In spite of the errors, the trends shown in Table (7.10) dramatically
display the influence of accounting for source term effects in evaluating the
interface variable values on the accuracy of resulting solutions. The same
problem was repeated using the explicit formulation for flux evaluation and
repeated SIMPLEC with SIP-BC or SIP-ACM to examine accuracy issues and CPU
requirements in comparison with the baseline code that incorporates Hybrid
differencing. The results for y . and CPU requirements are displayed in
Table 7.11.
TABLE 7.11 VALUES OF <|/m1n AND CPU REQUIREMENTS
FOR DRIVEN CAVITY PROBLEM

\.

Scheme
Hybrid

GritJ^^^

MW-SUDS-STI

.^
min

CPU

10x10

-0.0496

101

^min
4
0.0672

20x20

-0.0655

760

0.0879

CPU

LPU-SUDS-STI
,4
CPU
min

306

0.0734

285

1704

0.0954

1813

In the above and remaining tables to follow, CPU's are measured in seconds and
refer to a MASSCOMP Series 500 minicomputer.

,165-

Review of the above table again demonstrates clearly;that LP-SUDS-STI

solutions are more accurate than MW-SUDS-STI results which are in turn, more.,
accurate than Hybrid results. Associated with the increased accuracy of
MW-SUDS-STI or LP-SUDS-STI is a corresponding increase in total CPU requirements. A significant portion of this increase, apart fronrthe CPUrcost . ,.
associated with coefficient generation, (50 to 100 percent) 1s most likely due
to the use of the particular deferred correction procedure to account for .the
source term influence, primarily pressure for.the momentum conservation -equations. Naturally, such CPU requirements could be reduced substantially, if a
more implicit treatment of source term influences .were used. However, such an
implicit treatment of pressure source for interface flux evaluation may give
rise to practical solution difficulties within the framework of the segregated
solution approach adopted, i.e., decoupling of the pressure field, treatment
of boundary conditions for the additional terms in the continuity/pressure
(10)
equation, etc. Raw
using a similar approach formulated a co-located
solution procedure that overcomes most of the above problems. Use of the
deferred correction approach 1n the manner outlined above yielded converged
solutions without severe solution difficulties, albeit Increased CPU requirements.
.
.
The increased costs associated with the use of the above SUDS schemes
raises the issue of cost effectiveness of such schemes. One way to evaluate
cost effectiveness is to determine the relative costs of each scheme to
achieve a prescribed accuracy. For instance, the <ir
mm. value of -0.0734 is
obtained using LP-SUDS-STI on a 10x10 grid and requires approximately 300 CPU
seconds. On the same grid using a similar amount of CPU, the MW-SUDS-STI
result is -0.0672. It 1s estimated that MW-SUDS-STI would require a 15x15
grid and 700 CPU seconds to obtain a solution of similar accuracy as
LP-SUDS-STI on a 10x10 grid. Similarly, it is estimated that the Hybrid
scheme would require a 27x27 grid and 1800 seconds of CPU to obtain the
accuracy of the coarse grid LP-SUDS-STI result. With respect to the coarse
grid LP-SUDS-STI results, the storage and CPU requirements of the discretization schemes to obtain similar accuracy as presented in Table (7.12).

-166-

TABLE 7.12

STORAGE AND CPU RATIOS FOR EQUIVALENT ACCURACY

(DRIVEN CAVITY PROBLEM)

CPU Ratio

Scheme

Storage Ratio

Hybrid

7.3

6.0

MW-SUDS-STI

2.3

2.3

LP-SUDS-STI

1.0

1.0

For the flow over a rearward facing step, Reh =250 as shown in figure
(7.2), three uniform grids were used, 10x6, 20x12 and 40x24. From these numerical results, values of the normalized recirculation zone length L_K = x_/h
K
were determined to assess discretization accuracy. Table (7.13) displays the
L values obtained using an implicit interface variable determination with
and without source term influences, and SIMPLEC with SIP-BC. Also tabulated
are the Hybrid, QUDS and BSUDS2 reported by Syed, et al (5) and the experi(94)
mental results of Durst.
TABLE 7.13

VALUES OF LR FOR REARWARD FACING STEP PROBLEM

EXPERIMENTAL VALUE OF LR =6.3

Grid
10x6

20x12

40x24

Scheme
HYBRID
QUDS
BSUDS2

2.5
5.32
2.2

3.14
5.25
3.95

4.52
5.35
5.78

MW-SUDS
LP-SUDS

3.94
4.82

4.43
5.24

4.84
5.82

LP-SUDS-STI
LP-SUDS-STL

3.92
4.94

4.53
5.32

4.92
5.92

From the tabulated results it is noted that:

i)
MW-SUDS results are superior to Hybrid results,
ii)
LP-SUDS results are always superior to BSUDS2 results and
comparable or better than QUDS results on finer meshes.

-167-

iii)

In general, accounting for source term influence improves the

accuracy of MW-SUOS and LP-SUOS.
As in the previous test problem, application of the explicit interface
variable formulation together with repeated SIMPLEC and SIP-ACM in the baseline code, enables a comparative basis to be established for cost effectiveness of each scheme. These are summarized in Table (7.14).
TABLE 7.14

^\.Scheme
Grid
\^

VALUES OF LR AND CPU REQUIREMENTS FOR REARWARD

FACING STEP PROBLEM

Hybrid

LR

MW-SUDS-STI

CPU . LR

LP-SUDS-STI
CPU

CPU

LR

10x6

2.37

34

3.55

63

6.30

69

20x1 2

2.95

191

4.59

427

5.88

396

40x24

3.68

1265

5.33

2573

5.93

2495

A comparison of storage and CPU requirements for equivalent accuracy

based on the medium 20x12 grid LP-SUDS-STI results are also shown in Table
(7.15). The tabulated ratios are based on estimates that Hybrid would require
at least a 160x96 grid and 8.3 hrs of CPU, and MW-SUDS-STI an 80x48 grid and
4.3 hrs of CPU to approach the accuracy of the medium grid LP-SUDS-STI result.
TABLE 7.15

STORAGE AND CPU RATIOS FOR EQUIVALENT ACCURACY

FACING STEP PROBLEM)

Scheme

Storage Ratio

CPU Ratio

Hybrid

64

75

MW-SUDS-STI

16

39

LP-SUDS-STI

1,0

1.0

-168-

REARWARD

The final flow problem examined was that of turbulent coannular flow as
described in (5). For this case only the explicit interface variable formulation with repeated SIMPLEC and SIP-ACM was adopted to study accuracy issues
and cost effectiveness.
Regarding the application of SUDS schemes to predict the complex details
of a turbulent flow, the initial implementation of both MW-SUDS and LP-SUDS
included diffusion and source term influences in the equations for momentum,
turbulent kinetic energy, k, and turbulent kinetic energy dissipation, e, conservation. Although with appropriate treatment of k and e source term linearizations converged numerical solutions could be obtained, these solutions
contained regions where k and/or e were negative. Stability was maintained
only because negative k and e values were reset to zero. By the very nature
of equations for k and e, negative values for k and e can only arise through
the introduction of negative influences in the discretization schemes. After
a careful analysis of the discretization schemes it can be shown that negative
influences can arise in the following ways:
i)
As a result of the linear profile assumption made to relate A
to adjacent nodal values of 0, Section (6.3.1). Since this
assumption is inherent in LP-SUDS, this scheme is not appropriate for the discretization of equations for k or e. By design
MW-SUDS does not suffer from this problem.
ii)
As a result of including the influence of diffusion in the integration point equations. At present, the only solution to this
problem is not to include the diffusion term influence in the
integration point equation for k and c.
111')
As a result of including the influences of sources as currently
implemented. An alternative approach to including source term
influences may overcome this difficulty.
Guided by the results of the analysis given above, the baseline code was
modified so that only MW-SUDS without diffusion and source term influences was
used for k and e equations. Both LP-SUDS and MW-SUDS with diffusion and
source term influences were retained for the momentum equations. The modified
code was then used to obtain numerical solutions to the turbulent coannular
flow problem using Hybrid, MW-SUDS and LP-SUDS.
A comparison of the calculated centre-line axial velocity distribution
is shown in figure (7.19) for the coarse 21x20 grid. As indicated, the

-169-

1
0

COflRSE GRID
HYBRID
MU-SUDS & LP-SUDS

o
UJ

c/>
^

E
. 0
^

^
o 0
o

d 0
UJ

UJ

0.9

Figure 7.19

0.1

0.2

0.3

0.6 .
0.4
0.5
fl.XIflL POSITION

0.7

0.8

0.9

Centre-line Axial Velocity Distribution, Turbulent Coannular

Flow Problem, SUDS Schemes Using Coarse Grid

.1.0

FINE GRID

0.9

HYBRID

0.8

HU-SUDS 4 LP-SUDS

~ 0.6
o
o
0-5

0.4
cc

0.3

UJ

0.2
0.1
0.0

0.0

9.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

RXIflL DISTflNCE

W1667

Figure

0.1

7.20

Centre-line Axial Velocity Distribution, Turbulent Coannular

Flow Problem, SUDS Schemes Using Fine Grid

-170-

centre-line axial velocity distributions using MW-SUDS and LP-SUDS are identical. However, the Hybrid and SUDS results do not agree well in the initial
region. This discrepancy most likely arises because of the excessive smearing
of the Hybrid scheme.
Figure (7.20) shows the comparison of centre-line axial velocity distribution for the fine 38x38 grid. Again, there is no difference between MW-SUOS
and LP-SUDS results, and the Hybrid results are in closer agreement with the
SUDS results.
For completeness, information about coefficient updates and CPU requirements are summarized in Table (7.16).
TABLE 7.16 COEFFICIENT UPDATE AND CPU REQUIREMENTS FOR TURBULENT
COANNULAR FLOW PROBLEM

Grid

7.3.3

Scheme

Coefficient
Updates

Total CPU
Requirements
(hours)

21x20
21x20
21x20

Hybrid
MW-SUDS-STI
LP-SUDS-STI

109
137
182

0.5
1.5
2.0

38x38
38x38
38x38

Hybrid
MW-SUDS-STI
LP-SUDS-STI

421
457
451

8.87
20.6
20.7

Compact Implicit Discretization Schemes (CI)

The performance of several variants of CI schemes will be discussed

using the notation and terminology of Section (6.4), primarily in one or
two-dimensional scalar transport applications.
Derivative Compact Implicit Scheme (DCI)
DCI was applied to compute the following one-dimensional scalar transport problem including convection and diffusion:

-171-

d 0

0 < x<1
0 (x=0) = 0.0
0 (x=L) = 1.0
u >0

d0
dx

with the solution given by:

(7.4).

where Pe = uL/iv = Peclet number

v
Employing 31 uniformly spaced nodes and 5th order Pade1 approximations
for 0' and 0", OCI results for a range of grid Peclet numbers were obtained
and are displayed in figure (7.21). A close study reveals that for Pe = 1 the
solutions for 0 are bounded and physically correct. However, as the value of
Pe increases, the solution of 0 is no longer bounded with overshoots and
undershoots arising in the vicinity of the large gradient at x/L=l. In fact,
(82)
Ciment et al
have shown that the DCI scheme will give stable and
physically reasonable results for Pe < 4 / 15 = 2.1438.
Classical Operator Compact Implicit Scheme (COCI)
COCI scheme was also used to obtain solutions of the above onedimensional equation. The results show in figure (7.22) for various values of
Pe indicate for Pe<2 the correct behavior is obtained. However, as Pe is
increased beyond a value of 2 (corresponding to a decrease in diffusion) the
solution for 0 behaves as though more diffusion were introduced. In fact for
Pe =V^2~ the solution is identical to that for pure diffusion Pe = 0.0.
Increasing the value of Pe further results in even larger deviations from the
correct behavior. Finally, for Pe = 4.208, the COCI scheme generates an
entirely non-physical but bounded result, namely that all
the values of 0 in the domain equals the value at the downstream boundary.
Generalized Operator Compact Implicit Schemes (GOCI)
Using the above one-dimensional scalar transport equation GOCI results
were generated for a range of Peclet numbers and are displayed 1n figure
(7.23). Review of figure (7.23) clearly demonstrates the boundedness of

-172-

NODE

Figure 7.21

13

15

17

19 21

Compact Implicit, Block 3 x 3 , 5th Order Pade' BC.'s

ID Convection - Diffusion Problem, 31 Nodes

Figure 7.22

11

11

13

15 17 19 21

23

25 27 29 31

Standard OCI
ID Convection - Diffusion Problem, 31 Nodes

-173-

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3

X
0
D

= 1.0
= 2.0
= 4.0
= 00

0.1

i
11

0.0

Figure 7.23

i_ i
13

15 17
0.5
x

19

21

23

25

27

Generalized OCI
ID Convection - Diffusion Problem, 31 Nodes

5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37

9 11 13 15 17 19 21 23 25 27 29 31 33 35 37

y
(X =

Figure 7.24

0.5)

Convected Step Problem, GOCI, 0 = 45*

-174-

solutions for all Pe. Also, as expected from the inherent exponential
character of finite difference coefficients and the analytical solution, GOCI
solution is in excellent agreement with the latter.
Encouraged by the above results and notwithstanding the reservations
discussed in Section (6.4.5), GOCI Scheme was applied to predict the details
of two-dimensional angled transport of a boundary specified step profile of t
for various Pe's. The results shown in figures (7.24) to (7.26) for 8 = 45, .
33.7 and 18.4 respectively, all examine Peclet numbers of 250 and with
grids of 7x7, 19x19 and 37x37 nodes. Each figure displays the predicted profile for 0 along the centreline x/L = 0.5. The analytical solution for
Pe = is a step change in 0 from 0=0 to 0=1 at y/L = 0.5. The approximate analytical solution for Pe =250 case, valid for Pe >50, is not
plotted because it is graphically indistinguishable from the 37x37 node solutions for all flow angles considered.
At e=45, figure (7.24) the Pe = results for all grids are exact,
resolving the profile at y/L=0.5 to within the resolution of the grid. For
Pe =250, only the 7x7 grid prediction exhibits any distinguishable errors,
and even those are relatively small.
At 6=33.7, figure (7.25), the situation is somewhat different. For
Pe= , the predictions on all three grids exhibit significant overshoots and
undershoots at the outflow boundary where 0=1 is enforced. The magnitude of
the maximum overshoot (= 16 percent) does not reduce with grid refinement,
since the local grid Peclet Number remains infinite. The oscillatory behavior
tends to localize as the grid is refined, concentrating the overshoots and
undershoots close to the steep gradient. At Pe =250, the GOCI Scheme is
very well-behaved with only a small ( = 2 percent) overshoot arising on the 7x7
grid near the inflow boundary. For the 19x19 and 37x37 grids, the predictions
are comparable.
At e=18.4, figure (7.26), the results are similar to these described
for 8=33.7. The maximum overshoot on the downstream boundary for Pe=
remains at approximately 16 percent with behavior similar to that of 9=33.7,
as the grid is refined. For Pe =250, the 7x7 grid prediction has a maximum
overshoot of 4 percent. As the grid is refined to 19x19, the overshoots
disappear and the prediction is very close to that obtained using a 37x37 grid.

-175-

i i

357

i -i
9

I I I I Os->> A -A.jfc^iWi

lii

11 13 15 17 19 21 23 25 27 29 31 33 35 37

ii
3

i >vA- a Jfcfrfffi,

9 11 13 15 17 19 21 23 25 27 29 31 33 35 37

'

''

(x - 0.5)

(x = 0.5)

Figure 7.25

Convected Step Problem, 60CI, 0 = 33.7*

0.1

I I I I I I I I I
3

5 7 9 11 13 15 17 19 21 23 25 27 2931 33 35 37

9 11 13 15 17 19 21 23 25 27 29 31 33 35 37

y *

(x = 0.5)

Figure 7.26

57

(X 0.5)

Convected Step Problem, GOCI, 0 = 18.4'

-176-

control Volume Based Operator Compact Implicit Method of Exponential

Tyjg CVOCI
To demonstrate the applicability and determine the characteristics of
the CVOCI discussed in Section (6.4.3), the three scalar transport problems
chosen to assess the performance of SOU and SUDS schemes were adopted to provide a uniform comparison basis for all the schemes considered in this study.
For the uniform, angled flow convected step problem, the results of
CVOCI at the domain centre-line, x/L=0.5 are shown in figures (7.27). In general, the results indicate that:
i) for e=o, CVOCI generates no significant overshoots, even on the
coarse 5x5 grid;
ii) for 6=11.3 and 30.9, CVOCI generates a 10 percent undershoot
on the coarse 5x5 grid with the magnitude of the undershoot
diminishing with grid refinement;
iii) for 9=45, CVOCI generates only a mild 1 percent overshoot.
For the uniform angled flow past a unit source problem, the CVOCI pro(91)
files along the centre-line (y/L=0.5) and the exact results of Stubley
are shown in figures (7.28). The results for the 19x19 grid indicate that:
i) no overshoots or undershoots were observed with overall good
agreement between the CVOCI and exact results;
ii) for 8=0, the centre-line profile of 0 is overestimated by CVOCI;
iii) for 9=22.5, the centre-line profile of 0 is in excellent agreement with the exact solution;
iv) for 9=45, the centre-line profile of 0 is underestimated by
CVOCI.
For the Smith-Hutton
problem of curved flow of a diffused step
profile through a rectangular domain, the CVOCI profiles along the outlet
boundary are shown in figure (7.29). The results for this problem indicate
that
i) for Pe = 10 and 100, the coarse 20x10 grid CVOCI results are in
good agreement with the fine grid results presented by Smith and
Mutton;
ii) for Pe = 500 and 1000, the CVOCI results display a 10 percent
overshoot and a small 1 percent undershoot;
iii) for Pe=10 , the CVOCI results exhibit oscillatory behavior on
both sides of the steep gradient with errors as large as
17 percent arising.

-177-

1.0

0.75

1.0

9= OP, CVOCI
X/L - 0.5
PeL - 250

0.75

8 - 30.90, CVOCI
X/L-0.5
Pe L -250

c)

a)

0.5

0.5

0.25

0.25

0.0

0.0

1.0

1.0

8 - 450, CVOCI
X/L - 0.5
Pet - 250

0= 11.3, CVOCI
X/L = 0.5
PeL - 250
0.75

0.75

d)

b)

<f>

d>

0.5

0.25

0.25 -

0.0
0.0

0.5

0.0
0.5

1.0

y/L

Figure 7.27

Centre-Hne Profiles of 0 for Converted Step Problem, CVOCI

-178-

o
>
o
0)

o
a.
0)

u
a
o

1 -

' ^

in

*~. .

IP

. =

2
0

.'
'

"

'
'

'

o
.

/ ^

e
B //
V' <
,-^ X

":-

,-<-<
^- :"~~~

>

7/

&
><
"
r
"' ~~ '

i..
1

. J-

'

0)

I.
*->
c
0)

'v,.-:

o-

in ^
6 x
00
C\J

a>

O"

3
O>
rU-

SS22
M n n ^
11
-J s'S
<l> "isa

tf

i 2

o" *" *"

e ii n Q

.-

" " " ~rT*^

0 *

1 1
.m
*

0)
^

/'

Z
k
...**'
*\- ^-^*^"'
>'

O "? 0) O)

ii
o
^

i
in
o~

u>
o

P. i

o
d

.
o
r^

-179-

.i.
m
*"

5
i

t/1

i.
in
o

i.
in
ft

,o
o
o
d

-,9

.js/
O

0>

S
o

\>

V
q

q
f

in

OJ

m
6

q
6

</)

Q-

1/1
O)

., o

8
o

Ol

p
oi

a>

CM

in
6

o
6

q
N

e-

-180-

in
6

g"
o
*
0

o
2
O

E
0>

o
Q-

q
(M

q
T-

<.

6
P
6

1o

-&

t/>
d>

o
Q.

at
"^

a
o

c
o
o
0>
CM

0)

o>

m
o

-181-

o
6

7.4

CLOSURE AND SELECTION OF SCHEMES FOR THREE-DIMENSIONAL EVALUATION

This Section has presented the results of applying the previously

selected techniques to test problems in order to evaluate their performance in
enhancing convergence and efficiency as well as discretization accuracy. A
summary is now provided outlining the conclusions along with recommendations
for choices selected for three-dimensional evaluation. Convergence and efficiency enhancement aspects are discussed first.
Efficient numerical techniques for incompressible, viscous, recirculating flows should incorporate appropriate practices in the solution procedure
to deal with the following issues of concern:
i) techniques used to treat nonlinearities
ii) techniques used to treat coupling between variables, e.g.,
pressure and velocity, swirl, etc.
iii) techniques used to solve equations for single scalars, specifically, the equation for pressure correction that arises in general segregated solution procedures.
In the present study attention was directed primarily towards improved
techniques to effectively couple pressure and velocity and to solve efficiently the resulting pressure correction equation. Based on the results of
suitable numerical experiments designed to evaluate the performance of
repeated SIMPLEC together with SIP accelerated by CG, BC or ACM, the following
conclusions are reached:
i) The introduction of repeated SIMPLEC significantly enhances convergence, reducing both the number of coefficient updates and
CPU requirements. This is especially valuable for use with
coefficient updates of CPU intensive discretization schemes like
MW-SUDS and LP-SUDS.
ii) The introduction of an accelerated SIP to solve the pressure
correction equation can significantly reduce CPU requirements;
up to 40 percent for laminar flows and 20 percent for turbulent
flows. Concerning the choices for the acceleration technique
the following points are noted:
iii) Provided the computer storage is available, ACM provides the
most effective acceleration of SIP.

-182-

iv)

If it is known, a priori, that the solution for pressure varies

principally in one coordinate direction, then BC acceleration is
effective,
y) Conjugate Gradient acceleration exhibits poor convergence
behavior on fine grids.
Based on these conclusions it is recommended that to enhance the efficiency of numerical solution of three-dimensional viscous, recirculating
flows, SIP with ACM be used. Also, the improvement in efficiency provided by
ACM in two-dimensional flows is expected to be even more dramatic in threedimensional flows, where even on coarse grids, the number of unknowns is
large. Finally, for three-dimensional applications, the incremental storage
penalties associated with ACM are considerably smaller than they are in twodimensional applications.
Further consideration of repeated SIMPLEC algorithm, despite its
impressive performance in two dimensions, is outside the scope of the current
effort. Thus, PISO was adopted as the improved pressure ~ velocity coupling
algorithm in three dimensions.
Concerning improvements in discretization accuracy, the following
points are noted based on the results of numerical experiments presented above:
i) Second Order Upwind differencing solutions are more accurate
than solutions yielded by DOS, but can exhibit overshoots and
undershoots as well as smearing of gradients on coarse grids,
when the flow is at an angle to the grid. It is only on fine
meshes where the second order rate of convergence is approached
that the smearing and overshoots and undershoots diminish,
ii) Variants of Compact Implicit Schemes, specifically Control
Volume based Operator Compact Implicit Schemes, produce solutions that are substantially more accurate than UDS solutions,
but can exhibit relatively large overshoots and undershoots for
multi-dimensional flows when the flow is at an angle to the
grid. A probable cause for such non-physical solution features
is identified as due to errors that arise from an inadequate
treatment of flow curvature in the present implementation,
iii) Linear profile Skewed Upstream Differencing solutions have been
demonstrated to be more accurate than the above SOU or CVOCI
solutions with smaller overshoots and undershoots. Also for

-183-

laminar flow problems LP-SUOS has been shown to be at least 5

times more cost effective than UDS or its variants such as
Hybrid.
iv) Mass Weighted Skewed Upstream Differencing solutions have been
demonstrated to be free of overshoots and undershoots but not
free of numerical diffusion. However, the numerical diffusion
of MW-SUDS is considerably less than that of UDS. Consequently,
MW-SUDS solutions are, in general, more accurate than UDS solutions, and, for laminar flows, MW-SUDS has been shown to be at
least 2 times more cost effective than UDS.
v) The most accurate SUDS solutions are obtained when the Skewed
Upstream differencing is modified to account for the effects'of
diffusion and sources. However, for positive definite variables
like k and c, incorporating diffusion and source term influences
using the present implementation, together with the linear profile assumption is inappropriate. Thus, Mass Weighted skew " '
without diffusion and source term influences 1s adopted for
these equations with a corresponding degradation 1n performance,
:
Section (7.3.2).
,
i
j
Based on the above conclusions, the recommendations for Improved
accuracy in three-dimensional applications are now stated, guided by the following criteria:
'
1) Best accuracy on coarse grids: even with advanced computers,
many three-dimensional computations will still be limited to
:
grids which are, at best, coarse.
ii) Robustness and stability: for three-dimensional applications it
is important to ensure that numerical solutions are obtained.
Particular reference is made here to k and c equations,
iii) Efficiency: improvements in accuracy must be attained without
excessive additional computational requirements.
Three-dimensional computations that use UDS or Its variants such as
Hybrid should instead adopt LP/MW-SUDS for improved dlscretlizaton accuracy.
The use of LP/MW-SUDS will: 1) ensure robustness provided MW-SUDS is used
whenever the potential of minor overshoots and undershoots of LP-SUDS cannot
be tolerated; ii) significantly reduce or eliminate numerical diffusion,
thereby providing a considerable improvement in accuracy; and 111) provide
significant improvements in cost effectiveness. Naturally, modification of

-184-

skewed upstream differencing to include effects of diffusion and source terms

should be included in LP/MW-SUOS whenever possible. Future efforts should be
directed towards suitable incorporation of these effects in the k and e equations. Finally, if only one scheme is to be considered for prediction of
three-dimensional flows, the use of MW-SUDS, is recommended.

-185-

8.0

DISCUSSION OF THREE-DIMENSIONAL TEST CASE ,

In Section 7. various techniques for improving the computational efficiency of numerical methods for two-dimensional, vis'cous, recirculating flows
were tested and evaluated to assess convergence and accuracy issues. Based on
these results it was recommended that, for three-dimensional applications the
Strongly Implicit Procedure (SIP) with Additive Correction Multigrid (ACM),
and Linear Profile or Mass Weighted Skewed Upstream Differencing (MW-SUDS and
LP-SUDS) with corrections for diffusion and source term influences, when
appropriate, be used. Upon implementing these recommendations and testing the
resulting software, it was determined that several alternatives regarding
solver performance and additional considerations in the implementation SUDS
schemes were worthy of further study. These will now be discussed in detail.
8.1

EVALUATION OF TECHNIQUES FOR IMPROVING COMPUTATIONAL EFFICIENCY

The alternatives concerning the previously recommended solver choices

included the following:
i) using Incomplete Choleski (1C) instead of SIP to save computer
storage
ii) using 3x3x3 blocks for ACM instead of 2x2x2 blocks, again, to save
storage, Section (6.5.4)
iii) using Block Correction (BC) in conjunction with ACM to further
improve computational efficiency.
The merits and shortcomings of these alternatives are now assessed using
the following test problem in Cartesian coordinates
2
v 0 + S = 0 0 < x < 3 2

0 < y < 10

(8.1)

0 < z<8

where
30

TT = 0

at all boundaries anch

S (x,y,z) = 0 except S (0,5,z) = -1.0 and S (32,5,z) = 1.0. In equation (8.1)

2
v is the usual Laplacian and h is normal to the boundary. This problem

-186-

was chosen because the resulting discrete algebraic equations are similar to
the pressure correction equations of Section (6.1).
Using a Central Differencing Scheme (CDS) on a 32x10x8 grid, solutions
to the resulting algebraic equations were obtained using the various solution
techniques. Figures (8.1), (8.2) and (8.3) present the sum of the squares of
residuals versus computational effort for the various solution techniques. A
more detailed discussion of the results presented in each figure follows:

8.1.1

SIP vs 1C

:.

The recommendation to use SIP in conjunction with ACM is based on twodimensional results, where .the storage requirements of iterative solvers such
as SIP are not a primary concern. However, in three-dimensional applications
computational storage requirements can very often exceed the resources available. One approach to reducing the storage requirements is to use SIP with no
partial cancellation, which is algebraically equivalent to 1C. Storage requirements for various solution techniques are summarized in Table (8.1).
TABLE (8.1) STORAGE REQUIREMENTS OF ALTERNATIVE SOLUTION TECHNIQUES

Technique

Storage Words/Node

SIP
1C
SIP-ACM(2)
IC-ACM(2)
IC-ACM(3)

9.0
3.0
11.6
4.7
3.5
4.7
3.5

IC-BC-ACM(2)
IC-BC-ACM(31

As illustrated in figure (8.1), it is clear that the rate of convergence

of SIP with a near optimal partial cancellation is superior to the rate of
convergence of 1C. Even with the storage penalty it may be more advantageous
to use SIP over 1C. However, when ACM(2) (the (2) notation denotes blocks of
2x2x2 fine grid control volumes used) is used to accelerate SIP or 1C there is
little or no advantage to using SIP-ACM(2), especially if the cost of determining the optimal partial cancellation is taken into account. Also, the

-187-

1.0

Ql

I
O.OI

C.OOI

0.0001

10

20

30

ao

SO

CtHJISK)

Figure 8.1

Residual versus CPU for 1C, SIP, ACM(2)-IC and ACM(2)-SIP

OJ

O.OOOI

2O

30

40

SO

CPU fftcl

Figure 8.2

Residual versus CPU for 1C, ACM(2)-IC and ACM(3)-IC

-. 0.1

C.OI -

0.001 -

O.OOOI

20

30

to

so

CPU ISiCi

Figure 8.3

Residual versus CPU for 1C, BC-IC, ACM(2)-BC-IC and ACM(2)-BC-IC

-188-

results in figure and Table (8.1) indicate that it is more cost effective, in
terms of both computational storage and effort to use IC-ACM(2) .instead of
SIP-ACM(2).
8.1.2

ACMm VS ACM(3)

Another approach to reducing the storage requirements of the solution

techniques is to apply the Additive Correction Strategy to blocks consisting
of 3x3x3 fine grid control volumes, denoted by ACM(3), instead of blocks of
2x2x2 fine grid control volumes. As listed in Table (8.1) the storage requirements of ACM are reduced by 1.2 words/control volume by using blocks of
3x3x3 instead of blocks of 2x2x2. The results show in figure (8.2) indicate
that the reduced storage requirements of ACM(3) are offset by a lower rate of
convergence. Considering the minimal reduction in storage requirements and
lower rate of convergence of ACM(3), it is recommended that ACM(3) be used
only if the reduction in storage is necessary.
8.1.3

Block Correction (BC)

The results shown in figures (8.1) and (8.2) illustrate the dramatic
acceleration of convergence that results from using ACM. However, based on
two-dimensional results, an even more dramatic acceleration was expected.
This is especially true of problems, like the test problem in rectangular
domains and a solution having a significant (but not necessarily dominant)
one-dimensional component (in this case the x-direction). A more appropriate
acceleration technique for problems with a dominant one-dimensional component
is BC, Section (6.5.3). As illustrated in figure (8.3), BC provides an acceleration of 1C which is comparable to that of ACM(2) indicating that there is a
strong but not dominant one-dimensional component to the solution of the test
problem (if the one-dimensional component were dominant, IC-BC would converge considerably faster than IC-ACM(2)). As shown in figure (8.3), combining both ACM and BC to accelerate 1C, denoted by IC-BC-ACM(2), results in
convergence acceleration which is superior to either ACM or BC used alone.
Note also that there is no significant storage penalty associated with the use
of BC, Table (8.1).
These results indicate that for three-dimensional problems, the 1C base
solver does not always provide a sufficient amount of smoothing or relaxation
of the solution for ACM to be most effective. Combining BC with 1C provides

-189-

the necessary smoothing, particularly if there is a strong, but not

necessarily dominant, one-dimensional component to the solution.
8.1.4 Conclusions
These results, as well as others, indicate that IC-BC-ACM(2) is a solution procedure with a high rate of convergence over a wide range of problems
with a minimal storage penalty, and should significantly improve the computational efficiency of numerical methods used to predict three-dimensional
viscous, recirculating flows.
8.2

CONSIDERATIONS IN EXTENDING MW-SUDS AND LP-SUDS FORMULATION IN

THREE-DIMENSIONS
-,

The variants of SUDS Schemes, MW-SUDS and LP-SUDS, are designed to provide an accurate representation of the convective component of a general
transport problem within the framework of a finite volume discretization
approach. Using extensions of the two-dimensional ideas discussed .in :
Section (6.3), an approximation is required for the interface convective flux,
upon suitably integrating the parent transport equation over the finite
volume, i.e.,
C = / p v 0 dA
A

Equation

(6.4)

* ' .
.
where 0 is .the dependent variable, p is the density, v is the component of
velocity normal to the control volume face and A is the area of the control
volume face. In three-dimensions the above integral is evaluated using four
integration points on each control volume force, figure (8.4).
To express the variable values at the integration point in terms of
neighbouring nodal values, a discrete representation is required for the
following non-conservative differential transport equation in streamline
coordinates,
p

30
as

n
w

-.-..
Equation (6.39)

where V is the magnitude of the velocity, s is the streamwise coordinate and Q

represents all diffusive and source term influences. The different approaches
to obtaining representations of equation (6.39), as discussed in Section (6.3),
distinguish LP-SUDS from MW-SUDS.

-190-

 NODE POINT
x INTEGRATION POINT
Figure 8.4

Three-Dimensional Node Layout

on.
-f 101

00 If

0/0>-

000

Figure 8.5

too

Three-Dimensional Flux Element

OOlf

N1692

Figure 8.6

Flux Element with Control Volume Faces and Integration Points

-191-

To describe the manner in which the discrete representations of equation

(6.39) are arrived at in three dimensions, a systematic basis is required in
the following formulation, to describe the indexing associated with nodes,
integration points, flux elements, etc. This is introduced in figures (8.5)
and (8.6). It will be seen from figure (8.5) that the eight nodes surrounding
any integration point define the eight corners of a flux element, and the element so defined contain octants of eight different control volumes, figure
(8.6). Also each element contains twelve integration points with four integration points lying on each of the three planes that are coincident with control volume faces and intersect the flux element, figures (8.7), (8.8) and
(8.9).
In the interest of clarity, the triplet of binary indices illustrated in
figure (8.5) will be used to refer to a node or quantities associated with a
node. When reference is made to integration points or quantities associated
with the integration point, the symbols X, Y and Z along with the corresponding triplet of binary indices, shown in figures (8.6), (8.7), (8.8)
and (8.9) will be used.
<
8.2.1

Explicit LP-SUDS Integration Point Equation

To formulate a discrete representation of equation (6.39) for the integration point Oil lying on the control volume faces coincident with the
X plane, figure (8.7), the streamline is upwinded from the integration point
until it intersects a flux element side. The convective component of equation
(6.39) is then discretized, in a similar manner to the two-dimensional case,
as:
t

<u)

(8.2)

where 0 is the value of 0 at the point of intersection and L is the distance from the integration point to the intersection. In LP-SUDS, as discussed in Section (6.3), the value of 0 is determined from a bilinear interpolation of the node values of 0 lying on the intersected flux element side.
The discrete representation of the right-hand side of equation (6.39), Physical Advection Correction (PAC), is obtained from a trilinear interpolation of
the discrete nodal values of 0.

-192-

on
00 if-'-

0"'

000

Figure 8.7

/OO

Integration Points on Control Volumes Faces Coincident with

X-Planes

00/f

001

101

000

100

-1/10
"

000

Figure 8.8

/oo

Integration Points on Control Volume Faces Coincident with

Y-Planes

OOlr-

r'lOl

oig>--\
k000

/oo

N1681

Figure 8.9

Integration Points on Control Volume Faces Coincident with

Z-Planes

-193-

Rearranging, the discrete representation of equation (6.39) is expressed

as:
<8'3>

Examining the last term of equation (8.3), it is evident that some difficulties may arise when evaluating this term as V approaches zero. Fortunately,
as V approaches zero the grid Peclet number also becomes small (i.e., the Peclet number is much less than 2). Thus, the integration points representation
of 0 given in equation (8.3) can instead be evaluated using a trilinear interpolation of node values (cf. flux blending approaches),

'X011

^N
*
~*'NNP 0NP

<8/4>

where the N are the shape functions, borrowing from the finite element
approach, and the summation is made over the eight nodes of the flux element.
Since the Peclet number is low, the resulting CDS-like scheme will not generate negative downstream a coefficients as would occur for high Peclet numbers. To ensure that equation (8.4) is used instead of equation (8.3) for low
Peclet number applications, the integration point value is determined from a
weighted average of equations (8.3) and (8.4)

xon" (1"E
where B = Pe2/ (5 + Pe2) and is derived using the approximate exponential
weighting scheme of Raithby and Torrance
ensuring that no downwind negative a coefficients result.
It is seen from the above discussion that the explicit LP-SUOS formula
tion in three-dimensions follows readily from two dimensions, once an
appropriate set of indices is introduced.
8.2.2

Explicit HW-SUDS Integration Point Equations

In the process of extending the two-dimensional explicit MW-SUDS integration point equations to three dimensions, it became evident that the

-194-

extensions to three dimensions was not uniquely determined by analogy with the
two-dimensional algorithm. To overcome this, additional constraints had to be
imposed on the mass weighting logic. As a result, the following description
of explicit MW-SUOS is not a direct extension of the two-dimensional scheme.
However, in the two-dimensional limit the three-dimensional MW-SUOS presented
below is algebraically equivalent to two-dimensional MW-SUOS.
The primary difference between LP-SUDS and MW-SUDS lies in the evaluation of (u . In MW-SUDS the evaluation of 0u is chosen to ensure that the
resulting a coefficients are everywhere non-negative. To accomplish this the
evaluation of 0 is determined by logically deducing where the mass flows
crossing a control volume face originated. For instance, consider the case
where the mass flow crossing the X011 integration point control volume face is
leaving the Oil octant. The possible contributions to this mass flow are from
the mass flows through the two integration point faces, Y001 and Z010, as well
as the mass flows through the three octant faces intersecting at node Oil. To
ensure positive a coefficients it is assumed that any integration or node mass
flows leaving the octant cannot contribute to the X011 mass flow. Taking a
positive mass flow weighted average, 0 is approximated by:
F
U

YOOI "YOOI * F zoio gzoio * Fon *on

F

Y001 + F Z010 +F 011

"

(8 6)

'

where F represents the mass flow through the corresponding face of the octant
and the value of F is nonzero only if the mass flow is into the octant, and
where 0^001 and "zoio are estimates of tne corresponding integration
points. By similar positive mass weighting augments, the estimates of the integration points are given by:
F

xooi gon * Fzooo *zooo * Fooi *ooi

,
Y001

"

2010

"

X001

+F

ZOOO+F001

<8'7>

xoio Vii * F YOOO *YOOO * Foio Vio

F

X010+FYOOO+

-195-

010

(8 8)

'

where 0'
and 0'
are, again, estimates of integration* point values... . .
Also note that in equation ;(8.7) the node value> t , is used to approxi- ;
mate the; value.;of the integration point, tAUn..-,
This approximation is used
II
to ensure that an explicit relation for 0inm results. and that the resulting
/on\

AUUI

a coefficients are not negative. ^ ;' Similarly, t^- is used to approximate the value of <9X01Q-. Finally, positive mass weighting arguments ;and -.
approximations are used to obtain:
. -,: . .:,
F

ooo 'OOP-'*'- Fxooo gon * Fzooo

?nnn
2000

000 "OOP + FXOOO

FF

4- F
-i- F
000 + FXOOO + FYOOO

Combining equation (8.3) and equations (8.6) through (8.10) , 0 . . can be

Av I I
expressed in terms of the nodal values 0. , "nin' ''ooi and *000'
The evaluation of the PAC term of equation (8.3) 1s Identical to that
described for LP-SUDS except that the application of the PAC term may have to
be omitted if bounded solutions are to be guaranteed. Section (7.3.2). Therefore, for the solution of turbulent kinetic energy and dissipation PAC is not
incorporated in the formulation.
In the low Peclet number limit, MW-SUDS scheme outlined above is combined with a linear interpolation of nodes Oil and 1 1 1 , with the same weighting given in equation (8.5). For instance, for *xo,-,
'
"XOII (]-B> tN011 *011+ N lll "ill 3 +

B (

*u + 0-

Q)

(8.11)

Note that when PAC terms are omitted in MW-SUDS, the linear interpolation of
nodes in equation (8.11) is not required and B can be set to unity.
8.2.3

Flux Element Assembly

Using either the LP-SUOS or MW-SUDS approach, expressions explicitly

relating integration points to node points can be obtained for each integration point in the flux element. The influences of each integration point are
then assembled into the appropriate control volume flux balance equation by
substituting each integration point expression Into the appropriate control
volume face convective evaluation. The result is a control volume flux
balance equation which is expressed in terms of nodal points.
-196-

8.3 THREE-DIMENSIONAL TEST CASE

To illustrate the potential quantitative Impact of Incorporating the

above techniques to enhance discretization accuracy and solution convergence
in the baseline TEACH code. Section (7.2), a number of numerical predictions
were obtained. The test problem considered here refers to the flowfield characteristics of a typical three-dimensional turbulent flow 1n which a row of
jets are injected normal to the incoming flow in a duct with a rectangular
cross section. The problem details adopted (configurations, inlet velocity
and turbulence description, etc.) are similar to those examined in (5) with
the exception that the initial computations employed square cross-section jets
rather than circular. The case of circular jets along with comparison with
experimental data, provided by Khan
, are also presented for rigorous
quantitative evaluation.
.To illustrate the improvements in computational efficiency, the CPU
requirements of a Masscomp MSC Series 500 were measured to obtain predictions
using the baseline code and the modified code with ACM (2x2x2 blocks) in conjunction with IC-BC for the pressure correction updates of PISO. To achieve a
convergence of 0.005 using Hybrid differencing, the baseline code, on a 18x8x4
grid, required 80 iterations and 62 minutes of CPU. To achieve the same convergence with the same differencing and on the same grid, the use of IC-BC-ACM
reduced the iterations required to 39 and the CPU requirements to 29 minutes.
The results of this exercise indicate that, for the problem selected, the convergence of the baseline code is limited by the poor convergence characteristics of the TOMA used to solve the pressure correction equations and that the
IC-BC-ACM technique can dramatically improve overall convergence and computational efficiency.
To illustrate the impact of the improvements in accuracy, numerical predictions of the test problem were obtained using the baseline Hybrid differencing, MW-SUDS and LP-SUDS on 18x8x4 and 18x14x8 grids.
All of the calculations were performed using PISO with IC-BC-ACM used to
solve the pressure correction equations. The number of iterations required to
achieve a convergence level of 0.005 are presented 1n Table (8.2).

-197-

TABLE 8.2

NO. OF ITERATIONS FOR JET IN A CROSS FLOW PROBLEM

SCHEME

GRID

Hybrid
MW-SUDS
LP-SUDS
Hybrid
MW-SUDS
LP-SUDS

18x8x4
18x8x4
18x8x4
18x14x8

NO. OF ITERATIONS

18x14x8
18x14x8

39
46
53
61
74

i
'

82

The computational effort per iteration for the SUDS schemes is approximately 7 times that required by the Hybrid Scheme. This large increase is not
surprising considering that the SUDS schemes use four integration points per
control volume face compared to the one for Hybrid and that the calculations
required at each integration point are more complex for the SUDS schemes.
This increase in computational effort could be reduced significantly"with more
efficient coding techniques. Unfortunately, this would require a complete
overhaul of the existing TEACH Code.
Comparisons of calculated results for the axial velocity components at a
location 4 jet widths downstream of the jet centre-line are shown-in
figure (8.10). As expected, the Hybrid results exhibit what appears to be a
considerable amount of smearing with only minor improvements on the finer
grid. Using the modified code, the SUDS results exhibit considerably less
smearing with LP-SUDS producing results with the least amount of smearing.
These results Indicate that the accuracy of turbulent flows can be improved
using the modified SUDS schemes and, considering the excessive smearing
exhibited by Hybrid, even on finer grids, the additional cost per iteration of
the modified SUDS schemes may very well be offset by the improved accuracy of
the results.
'
, .- .?..
Considering the performance of LP-SUDS in the light of experimental
data, figures (8.11), (8.12) and (8.13) display axial velocity profiles at
locations 4 and 6 jet widths downstream of the jet centre-line using grids of
34x10x15 and 40x20x17 respectively. The grids used in these predictions
approximate the circular cross-section of the jets as well as the axial domain

-198-

4.0
GRID-- I8*&*4
- 10.8 mis

3.0

a)

Y/D 2.0

- HYBRID
* MW-SUDS
+ LP-SUDS

1.0

0.0
0.25

. TN
0.5

0.75

1.0

1.25

4.0
GR/D--18x/4x8
= 10.8 m/s

b)

3.0

Y/D 2.0

HYBRID
MW-SUDS
LP-SUDS

1.0

0.0
0.25

Figure 8.10

\
0.5

0.75

1.0

1.25

1.5

Axial Velocity Profiles 4 Jet Widths Downstream of the Jet

Centre-line

-199-

c/i
at at

c
a *- i
c o i M8

(A

a. t_

o
-a
at > c iM -> at &.

-J f- J CD

a.

3
o
^~

S
to
d

o
o -> 10
r- at
at >~j o>

O > E~

O i/>
U- fO t. 3

o f- u.
X

P~

O r-

O
II X
</> T3 II
r- O)
Q O

4. u ea *s. CM
<O 3 *V N

<
cc

a. i/i x
x
E <o "O
o at 4-> c o
<_> X to 10 <

CNJ

oo

S
ri

8
fj

a/A

.^

S 8-

S
o

0)

8
d

o>

NOiioarNi WOHJ aoNvisia aaznvwuoN

0) 0)

8 *

C O I to fc. 0) O

a. _

-M -O

O) >> C ^J +J O) S_

3 O ->
U r- tt)

>

o
UJ

4- IB 4- 3
Of- LlX
O

C <

Q. -ui X

a> 4->-o
c

ri

a/A '

o
n

s
<M

I6

NOiioarNi WOHJ aoNvisia aaznvwuoN

-200-

o
o X <o <o

I
Z

GO

o
in

II

t> -o II
i- aj o in
i- i- o ^ p<o 3 "^ N

L.
3
O>

DATA
=10.8

4.00

m/sec

Q
^
llf

3.50
BSUOS (Syed)
P-SUDS
3.0

o
HI

O
cc
LL

2.50

2.00

LU

O
1.50

Q
HI
N

1.00

CC
O

0.50

0.00
0.00
N1666

Figure 8.13

0.40

0.80

1.20

1.60

2.00

NORMALIZED AXIAL VELOCITY, U/Uoo

Comparison of Calculated and Measured Axial Velocity Profiles
at x/D = 6 From Jet Centre-Hne and z/0 = 0 Using a Grid
of 40 x 20 x 17

-201-

width in a manner identical to that reported in (5), but otherwise employ a

uniform grid distribution in the radial direction. Also, plotted on the above
figures are the corresponding experimental data provided by Khan
and the
predictions obtained by Syed, et al
using bounded skewed differencing.
The initial jet velocity profile used for the predictions incorporate
the experimentally measured centre-line profile,
but otherwise use a mass
conserving similarity criteria to provide the variation in the tangential and
axial width of.the jet. The .inlet profile for the approach flow assumed a top
hat profile due to unavailability of experimental data. Use of such an idealized profile, considering the proximity of the jets to the inlet, along with
the assumed turbulence intensity levels (1 percent for both flows) are serious
issues of concern for rigorous quantitative evaluation. However, figure (8.11)
revelas that the basic profile shape is predicted well with some radial dis-:,
placement in the peak minimum velocity location. This latter feature is characteristic of profiles generated by both grids and would tend to suggest needA
for a closer examination of the various inlet profiles used. Also, the,
coarseness of the grid in the vicinity of the jets, as well as the inappropriateness of the k-e model to model severely anisotropic turbulence fields,
(97)
'<*
Crabb,
might have contributed to the observed behavior. Even though the ,
bounded skewed differencing profiles agree with the experimentally measured
values better at certain locations, both near the jet inlet and the roof, '*
significant discrepancies are apparent due to the coarseness of the grid ,-.
employed. In general, LP-SUOS profiles are less corrupt by numerical
"~
diffusion but tend to overpredict experimental values, possibly due to poor -,
.i
modelling of turbulence and inlet profiles.
T
Figures (8.12) and (8.13) display the axial velocity profiles obtained
at 4 and 6 jet widths of the jet centre-line using a grid of 40x27x17. The '
solution difficulties reported by Syed, et al
in determining an appropriate local blending factor for use with bounded skewed differencing (corresponding results are also displayed) were absent from LP-SUDS predictions. It
is seen from the above figures that LP-SUDS produces improved profiles that
agree well with the experimental values in basic shape, albeit radially
shifted, .with the numerical values of the peak minimum velocities in much
closer agreement with experiments than bounded skew predictions. However, the
two schemes in this case yield profiles that agree closely over a larger portion of the flow field (hence significant improvements over Hybrid differencing) in spite of the large discrepancies observed in the vicinity of the jet

-202-

inlet. The reservations expressed in relation to figure (8.11) are also valid
for these figures in quantitatively evaluating the performance of LP-SUDS with
experimental data.
Although significant improvements are observed in LP-SUDS prediction
with grid refinement from a study of figures (8.11) and (8.12), it is not
assured that the solution yielded by a grid of 40x27x17 (still coarse for a
problem of this kind) is grid independent. Local grid refinement near the jet
inlet might be required to capture the relevant flow details. However, the
computational cost associated with obtaining grid independent solutions during
the present effort was prohibitive.
8.4 CLOSURE

This Section has developed and assessed the previously selected SUDS
schemes, as well as various accelerated schemes appropriate for pressure correction equation of segregated solution algorithms for three-dimensional
applications. For modelling the jet in a cross flow problem, adopted as the
relevant test case, significant improvements in overall convergence and computational efficiency were provided by the use of IC-BC-ACM, while the variants
of SUDS yielded solutions that exhibit considerably less smearing than those
of Hybrid differencing. A number of questions pertaining to inlet profile
specification, along with the previously reported shortcomings of k-c to model
the details of the anisotopic turbulence field generated by a jet in a cross
flow, were raised for a quantitative assessment. It was also concluded that,
in spite of the significant additional cost associated with the use of improved SUDS schemes in three-dimensional applications, their cost effective
use is still recommended on grounds of improved accuracy, as Hybrid solutions
are plagued with excessive smearing.

-203-

9.0
9.1

CLOSURE

SUMMARY AND DISCUSSHJN

-- . >. i.;r

; ,-

,-...-. r- >j

Numerical techniques to predict incompressible, turbulent, viscous,

recirculating flows frequently encountered in gas turbine engine components
have undergone rapid development within the last decade. However, some fundamental problems still exist that limit the performance of such methods in
routine engineering computations. Such issues refer to discretization
inaccuracies -introduced when the parent differential transport equations are
replaced by discrete algebraic equations and the computer resources' required
to so.lve these discrete equations. This study was motivated by the need1 to
address both of these problems in a comprehensive manner.
.
The effort was initiated by identifying the desirable attributes that ...
optimum discretization schemes and cost effective solution methodologies
shou.ld possess to provide a unified assessment criteria for subsequent
quantitative/qualitative evaluations. As adopted in the present study, these
criteria comprise guidelines to define critical issues of accuracy, stability
(robustness), efficiency, storage requirements and ease of implementation in a
three-dimensional code structured in TEACH methodology. The initial evaluation of more than ten potential techniques considered was primarily based on
examination of accuracy and linear stability of the resulting difference equations via evaluation of the properties of the coefficient matrix, Taylor series analysis and existing heuristic stability analyses for iterative solvers >
commonly used in segregated solution procedures.
* .
Upon completion of this initial evaluation four of-the most promising techniques were incorporated in a variant of 2D-TEACH code for further quantitative evaluation using carefully selected test problems. The particular
schemes selected address issues of both discretization accuracy and convergence enhancement. The latter concern in this study specifically refers to the
solution techniques adopted for the pressure/correction equation(s) generated
by general segregated solution algorithms for incompressible flows, although
similar practices could be equally employed in solving for momentum and other
scalars. However, the pressure/correction equation which displays a symmetric

-204-

structure with properties similar to those for a diffusion process, accounts

for a major portion of the solution cost associated with segregated solution
procedures.
Effective solution of the pressure/correction equation in this study was
accomplished by adopting approximate factorization techniques, including the
Incomplete Choleski (1C) and Stone's Strongly Implicit Procedure (SIP), in
place of the Alternating Direction Line Gauss-Seidel (ALGS) commonly used.
However, the high initial convergence rate displayed by 1C and SIP in test
problems due to removal of high frequency error components was severed, as the
asymptotic convergence rate is usually dictated by removal of low frequency
error components which are not diminished effectively by such base solvers.
To overcome this difficulty and to simultaneously address the complete error
spectrum, a series of accelerators including Conjugate Gradient (CG), Block
Correction (BC) and Additive Correction Multigrid (ACM) were considered for
use with 1C and SIP.
.
The performance of these accelerated solvers were evaluated in a variety
of test problems where the behavior of IC-CG, as measured by residual reduction, was shown to be erratic for many iterations. For fine grid performances
of interest here, the number of iterations required by IC-CG to reach the
point of precipitous residual reduction were found to be prohibitive, as each
iteration establishes an orthogonal base vector of the solution in a manner
that minimizes a prescribed norm of the error. For the same single scalar
problems, accelerations of both SIP and 1C by ACM resulted in exceptionally
improved (comparable) convergence rates, while BC acceleration of these
respective solvers yielded improvements in performance that range from
dramatic to moderate. The cost reduction factor associated with the introduction of ACM accelerated SIP for pressure/correction solution in a segregated
solution procedure was found to be up to 40 percent for laminar flows and 20
percent for turbulent flows.
For incompressible flow computations using segregated solution methodologies there also exists substantial additional cost associated with the appropriate coupling of pressure and velocity fields to yield the required zero
divergence for mass. In this study (for two-dimensional applications)
repeated SIMPLEC was adopted to enhance convergence by reducing both the number of coefficient updates and CPU requirements. The introduction of repeated
SIMPLEC was shown to have the largest impact on the high cost of coefficient
updates of CPU intensive schemes like MW-SUDS and LP-SUDS.

-205-

Concering the discretization schemes selected for quantitative -evaluation using a finite volume method, a unified framework was devised to delineate the nature of the approximations introduced by the various treatment of
convective fluxes at the cell interfaces. This involved a study of the
separate components of the discretization error (profile and operator errors)
in a manner indicated by Stubley (74) and concluded that a low truncation
error profile used for convective fluxes might not necessarily result -in small
solution error, because the finite difference operator does not embody the
correct physical influence. Furthermore, poor physical influence schemes,
that fail 'to correctly incorporate the influences of convection, diffusion and
source terms in the discretization procedure, were shown to invariably result ~
in algebraic equations that are difficult to solve using iterative methods and
possibilities for physically unbounded solutions exist.
Viewed in this light, the schemes selected, notably Second Order Upwind
(SOU) differencing and appropriate variants of Compact Implicit schemes (CI),
are seen to incorporate increasing physical corrections in their respective
grid upstream formulations, over the conventional Upwind differencing (UDS).
However, due to the possibility of generating negative coefficients and the
inappropriateness of the strictly one-dimensional boundedness in criteria in
multi dimensions, the solutions generated,by these schemes suffered from overshoots and undershoots for some test problems, especially when the flow was atr
an angle to the grid. Specifically, SOU solutions were found to be more r
accurate than UDS solutions with accompanying overshoots and undershoots as
well as smearing of gradients on coarse grids, when appreciable streamline
inclination to the grid was present. In fact, only on fine meshes where the
second order rate of convergence for SOU was approached that the smearing and
overshoots and undershoots diminished. Variants of CI schemes, particularly
CVOCI, that implicitly relate the nodal values of the variable and the
operator yielded significantly more accurate solutions than UDS. However, '"
using the current implementation, probabilities of generating relatively large
overshoots and undershoots for multi-dimensional flows still exist in the
presence of streamline skewness to the grid for high Peclet numbers.
Skewed Upstream Differencing Schemes (SUDS) eliminate the need for the
convective component of the PAC term by a suitable transformation ignored in
Grid Upstream Schemes, but otherwise incorporate the correct physical

-206-

Influences of diffusion and source terms in their formulation. The two variants examined in this study are distinguished merely by the definition of the
upstream value of the variable which, in turn, determines whether the unbounded, but accurate LP-SUOS or physically bounded, but less accurate MW-SUDS
results. Both LP-SUOS and MW-SUDS using implicit and explicit integration
point values were evaluated in a variety of test problems in which different
balances were dominant to examine their respective accuracy, stability and
cost effectiveness characteristics. It was concluded that the Physical
Advection Correction (PAC) to LP-SUDS leads to a full second order scheme that
was shown to be robust, although the solution cannot be guaranteed a priori to
be physically bounded. MW-SUDS, in turn, provided the best alternative, among
the schemes examined, for problems where physical bounding is mandatory.
The above quantitative two-dimensional evaluations also identified
LP/MW-SUDS coupled with SIP accelerated by ACM as deserving further evaluation
on three-dimensions, based on the practical criteria of best accuracy on
coarse grids, robustness and stability and efficiency. Subsequently, these
schemes, with appropriate modifications, were incorporated in a variant of
3D-TEACH code and were further evaluated regarding issues of accuracy'and
convergence enhancement characteristics in modelling of a jet in cross flow.
Dramatic improvements in overall convergence and computational efficiency were
realized for this problem regarding the solution of the pressure correction
equations using IC-BC-ACM, while use of SUDS schemes yielded considerably
reduced smearing of the solution, hence justifying the additional cost per
iteration associated with their use.
It can be said in conclusion that for the problems examined, this study
has clearly demonstrated that appropriate solution techniques for incompressible, turbulent, viscous recirculating flows in current use benefit substantially from the introduction of the convergence enhancement techniques adopted
here (up to 40 percent). Furthermore, the various improved discretization
schemes considered, specifically improved SUDS with PAC, usually provide a
significant improvement in accuracy and hence cost effectiveness relative to
most of the schemes currently used.
This study has also identified various areas where further research
might prove beneficial. These will now be discussed briefly.

-207-

9.2

RECOMMENDATIONS FOR FUTURE RESEARCH

'

. ,

i) As the results of this study are based on a limited and carefully

selected set of problems, the performance of the techniques considered should
be evaluated on a broader scope of test problems that emphasize relevant numerical aspects, before more general conclusions can be reached regarding their
utility.
ii) The current treatment of source terms in the integration point
equations of SUDS is explicit. For the discretized momentum equations that
incorporate pressure gradients as additional sources in the determination of
integration point variable, such a practice when employed in a segregated
solution procedure yields the usual pressure/correction equation. While this
deferred correction approach eventually provides the required solution at convergence, it significantly increases the number of coefficient updates. What
is required is an implicit treatment of pressure gradient sources for momentum
especially, coupled with the appropriate numerical techniques to deal with the
potentially troublesome pressure/correction equation.
iii) The current discretization for the positive definite variables
like k and e, adopts MW-SUDS without considering the influences of source arid
diffusion in the integration point equations. Such a numerical treatment for
primarily source and diffusion dominated k and c (in most regions of the flow)
may not be appropriate for flows driven by the details of the turbulence field
(aside from the physical modelling issues of turbulence). Thus, alternative
formulations that properly incorporate the predominant source term influences
have to be evaluated. A related concern pertains to the current implementation of MW-SUDS which has been shown to be frequently prone to excessive numerical diffusion. The implications of adopting a MW-SUDS formulation,
suggested by Huget,
that simultaneously attenuates the inherent numerical diffusion and eliminates the possibility of unbounded solutions should
be examined.
iv) This study has adopted the use of a nine point Alternating
Direction Line Gauss-Seidel procedure to solve the linear algebraic equations
generated by LP-SUDS (and SOU), even in the presence of mildly negative
coefficients, without any serious solution difficulties. However, solution
stability and convergence via such solvers is not strictly assured. To
circumvent potential problems associated with these and alternative discretization schemes, that introduce computational molecules incompatible with ADI,

-208-

development and use of appropriate matrix solvers should be seriously

considered.
v) Current solution algorithms for incompressible flows invariably
employ a staggered grid arrangement for variables to prevent the decoupling of
the pressure field. The PAC approach used here for improved accuracy, specifically the correction of momentum fluxes via pressure gradient source correction, however, does provide naturally the coupling between the pressure and
velocity fields to eliminate the need for such cumbersome and unwieldy prac(10)
tices. Raw
has recently incorporated similar ideas in his co-located
finite element algorithm. Furthermore, using such a co-located solution procedure, implementation and use of adaptive gridding via moving or stationary
adaptive grids to resolve fronts better, becomes manageable and should be
examined.
vi) The particular formulations of SOU and variants of CI examined in
this study incorporate various Grid Upstream practices resulting in inaccuracies and solution difficulties in the presence of appreciable streamline skew-;
ness to the grid for high Peclet numbers. Considering the high accuracy
provided by GOCI schemes (CVOCI in particular), albeit accompanied by nonphysical oscillations, a natural multi-dimensional extension for such schemes
that does not violate the one-dimensional boundedness criteria should consider
a "skewed" formulation. In this manner it is possible to retain strict
implicitness between the nodal values of the variable and the operator to
yield improved accuracy, while eliminating the non-physical solution wiggles.
vii) The above formulation for GOCI will also require development and
use of appropriate cost effective solvers in multi dimensions for the scheme
to be routinely used in engineering computations. Such solvers must recognize
the implications inherent in the equations, while using practices to substantially reduce the solution cost associated with block coupled systems.
viii) The significant reduction in cost provided by the introduction of
repeated SIMPLEC in the current segregated solution procedure was demonstrated
for two-dimensional problems using CPU intensive schemes like LP/MW-SUOS. In
practical three-dimensional applications, where the coefficient update costs
associated with such schemes become almost prohibitive for routine use,
repeated SIMPLEC should be incorporated in the solution algorithm. Furthermore, most current segregated solution algorithms including SIMPLEC degrade

-209-

(22)
severely in performance with grid refinement.
Thus, development and use
of "grid-insensitive" segregated solution procedures in the same philosophy as
SIMPLEX /?2) should provide substantial savings for appropriate problems.
ix) While this study emphasized the relevant details of the predominant
pressure-velocity coupling, appropriate for "simple" incompressible flows, by
a careful study of the various approximations employed, additional and/or
alternative couplings might significantly influence the solution characteristics for some problems. For instance, the nature of the couplings provided by
swirl (not considered in this study) as well as turbulence variables should be
examined and appropriate practices be developed.
x) The turbulent test cases examined in this study introduce additional
considerations in quantitatively evaluating the performance of discretization
schemes. Appropriateness of the physical modelling assumptions implied by the
particular turbulence model and availability of benchmark data for initial and
boundary condition specification, as well as quantitative assessment, significantly influence the nature of the solutions. For such problems improved predictions are generally obtained by solution procedures that incorporate
appropriate physical and computational models, in addition to improved
discretization techniques.
xi) Finally, a deeper understanding and appreciation of accurate discretization techniques and solution algorithms should be developed. This
study has only presented a phase of progress in a complex but steadily evolving subject. For example, future numerical techniques for the problems considered here might adopt spectral iteration techniques in conjunction with a^
lower order scheme (like SUDS) to achieve higher accuracy via a deferred cor(53)
. rector approach.
Also fully coupled numerical schemes/solution
(20)
algorithms
might be appropriately formulated to yield practical solutions for engineering problems.

-210-

:..'

' '*'

'-.

10.0

REFERENCES

1.

Patankar, S.V., Numerical Heat Transfer, Hemisphere, Washington, D.C.,

1980.

2.

Kenworthy, M.J., Correa, S.M. and Burrus, O.L., "Aerothermal Modelling

Program, Phase I-Final Report", NASA CR-168296, 1983.

3.

Sturgess, G.J. "Aerothermal Modelling Program, Phase I-Final Report",

NASA CR-168202, 1983.

4.

Srinivasan, R., Reynolds, R., Ball, I., Berry, R., Johnson, K. and
Mongia, H., "Aerothermal Modelling Program, Phase I-Final Report",
NASA-CR-168243, 1983.

5.

Syed, S.A., Chiappetta, L.M. and Gosman, A.D. "Error Reduction Progran,
Final Report", NASA-CR 174776, 1985.

Issa, R.I., "Solution of Implicitly Discretized Fluid Flow Equations by

Operator Splitting", Rept. FS/82/15, Imperial College, London, 1982.

7.

Launder, B.E. and Spalding, D.B., Mathematical Models of Turbulence,

Academic Press, London, 1972.

8.

Launder, B.E. and Spalding O.B., "The Numerical Computation of Turbulent

Flows", Comp. Meth. in Appl. Mech. and Eng. 3, 269, 1974.

9.

Harlow, F.H. and Welch, J.E., "Numerical Calculation of Time-Dependent

Viscous Incompressible Flow of Fluid with Free Surface", Phys. Fluids,
8, 12, 2182, 1965.

10.

Raw, M.J., "A New Control Volume Based Finite Element Procedure for the
Numerical Solution of the Fluid Flow and Scalar Transport Equaitons",
Ph.D. Thesis, University of Waterloo, 1985.

11.

Spalding, D.B., "A Novel Finite Difference Formulation for Differential

Expressions Involving Both First and Second Derivatives", Int. J. Num.
Meth. Eng., 4. 551, 1972.

12.

Raithby, G.D., "Skew Upstream Differencing Schemes for Problems

Involving Fluid Flow", Comp. Meth. Appl. Mech. Eng., 9, 153, 1976.

13.

Zedan, M. and Schneider, G.E., "Investigation into the Simultaneous

Variable Solution of Velocity and Pressure in Incompressible Fluid Flow
Problems", AIAA Paper No. 83-1519, presented at the AIAA 18th
Thermophysics Conference, Montreal, Quebec, Canada, 1983.

14.

Van Doormaal, J.P., "Numerical Methods for the Solution of

Incompressible and Compressible Fluid Flows", Ph.D. Thesis, University
of Waterloo, 1985.

-211-

15.

Van Doormaal, J.P. and Raithby, 6.D., "Enhancements of the SIMPLE Method
for Predicting Incompressible Fluid Flows", Numer. Heat Transfer, 7,
147, 1984.

16.

Gosman, A.D. and Lai, K.Y., "Finite Difference and Other Approximations
for the Transport and Navier Stokes Equations", FS/82/18, Imperial
College, London, 1982.
,

17.

Forsyhte, G.E. and Wason, W., Finite Difference Methods for Partial
Differential Equations, Wiley, New York, 1960.

18.

Roache, P.J., Computational Fluid Dynamics, Hermosa, Albuquerque, N.M.,

.1972.

19.

Van Doormaal, J.P., Turan, A. and Raithby, G.O., "Evaluation of New

Techniques for the Calculation of Internal Recirculating Flows", AIAA
Paper No. 87-0059, presented at the AIAA 25th Aerospace Sciences
Meeting, Reno, NV, 1987.

20.

Vanka, S.P. and Leaf, G.K. "Fully Coupled Solution of Pressure-Linked

Fluid Flow Equations", ANL-83-73, Argonne National Laboratory, Argonne,
IL, 1983.

21.

Patankar, S.V., "Efficient Numerical Techniques for Complex Fluid

Flows", NASA CP-2444, 141, 1986.

22.

Van Doormaal, J.P. and Raithby, G.D., "An Evaluation of the Segregated
Approach for Predicting Incompressible Fluid Flows", ASME Paper 85-HT-9,
presented at the National Heat Transfer Conference, Denver, CO., 1985. ,

23.

Van Doormaal, J.P., Hutchinson, B.R. and Turan, A., "An Evaluation of
Techniques Used to Accelerate Segregated Methods for Predicting Viscous
Fluid Flow", AIAA paper No. 86-1653, presented at the AIAA/ASME/SAE/ASEE
22nd Joint Propulsion Conference, Huntsville, AL, 1986.

24.

Schumann, U. and Sweet, R.A., "A Direct Method for the solution of
Poisson's Equation with Neumann Boundary Conditions on a Staggered Grid
of Arbitrary Size", J. Comp. Phys., 20, 171, 1976.
,

25.

Hageman, L.A., and Young, D.M., Applied Iterative Methods, Academic

Press, New York, 1981.

26.

Stone, H.L., "Iterative Solution of Implicit Approximations of

Multi-Dimensional Partial Differential Equations", SIAM J. Numerical
Analysis, 5, 530, 1968.

27.

Schneider, G.E. and Zedan, M., "A Coupled Modified Strongly Implicit
Procedure for the Numerical Solution of Coupled Continuum Problems",
AIAA Paper No. 84-1743, presented at the AIAA 19'th Thermophysics
Conference, Snowmass, CO, 1984.

28.

Atias, M., Wolfshtein, M. and Israeli, M., "Efficiency of Navier Stokes

Solvers", AIAA J., 15, 263, 1977.

-212-

29

Gupta, M.M. and Manohar, R.P., "A Critique of a Second-Order Upwind

Scheme for Viscous Flow Problems", AIAA J., 16, 759, 1978.

30.

Huget, R.G., "The Evaluation and Development of Approximation Schemes

for the Finite Volume Method", Ph.D. Thesis, University of Waterloo,
1985.

31.

Orszag, S.A. and Israeli, M., "Numerical Simulation of Viscous

Incompressible Flows", Annual Review of Fluid Mechanics, 6, 281, 1974.

32.

Orszag, S.A., "Numerical Simulation of Incompressible Flows Within

Simple Boundaries: Accuracy", J. Fluid Mech., 49, 1, 75, 1971.

33.

Arakawa, A., "Numerical Simulation of Large Scale Atmospheric Motions",

Numerical Solution of Field Problems in Continuum Physics, Proc.
SIAM-AMS, Vol. 2, 24-40, Providence AMS, 1970.

34.

Phillips, N.A., "An Example of Non-Linear Computational Instability",

The Atmosphere and the Sea in Motion, 501, New York, Rockefeller
Institute Press, 1959.

35.

Rubin, S.G. and Khosla, P.K., "Higher Order Numerical Methods Derived
from Three Point Polynominal Interpolation", General Applied Science
Labs. Inc., NASA CR-2735, 1976.

36.

Ciment, M., Leventhal, S.H. and Weinberg, B.C., "The Operator Compact
Implicit Method for Parabolic Equations", Naval Surface Weapons Centre,
NSWC/WOL/TR-77-29, 1977.

37.

Berger, A.E., Solomon, J.M., Ciment, M., Leventhal, S.H. and Weinberg,
B.C., "Generalized OCI Schemes for Boundary Layer Problems", Math. Comp.
35, 151, 695, 1980.

38.

Von Neumann, J. and Richtmeyer, R.D., "A Method for Numerical

Calculation of Hydrodynamic Shocks", J of Applied Physics, 21, 232, 1950.

39.

MacCormack, R.W., "A Numerical Method for Solving the Equations of

Compressible Viscous Flow", AIAA Paper No. 81-0110, 19'th Aerospace
Sciences Meeting, Reno, NV, 1981.

40.

Jameson, A., Schmidt, W. and Turkel, E., "Numerical Solutions of the

Euler Equations by Finite Volume Methods Using Runge-Kutta Time Stepping
Schemes", AIAA Paper NO. 81-1259, 14'th Fluid and Plasma Dynamics
Conference, Palo Alto, CA, 1981.

41.

Beam, R.M. and Warming, R.F., "An Implicit Finite Difference Algorithm
for Hyperbolic Systems in Conservation-Law Form", J. Comp. Phys. 22, 87,
1976.

-213-

42.

MacCormack, R.W. and Baldwin, B.S., "A Numerical Method for Solving the
Navier-Stokes Equations with Application to Shock-Boundary Layer
Interactions", AIAA Paper No. 75-1, Pasadena, CA, 1975.

43.

Pulliam, T.H., "Artificial Dissipation Models for the Euler Equations",

AIAA paper No. 85-0438, 23'rd Aerospace Sciences Meeting, Reno, NV, 1985.

44.

Kwak, D. and Chang, J., "An Incompressible Navier-Stokes Flow Solver in

Three-Dimensional Curvilinear Coordinate System Using Primitive'
Variables", AIAA Paper No. 84-0253, 22'nd Aerospace Sciences Meeting
Reno, NV, 1984.

45.

Chorin, A.J., "A Numerical Method for Solving Incompressible Viscous

Flow Problems", J. Comp. Phys. 2, 12, 1967.

46.

Boris, J.P. and Book, D.L., "Flux-Corrected Transport, I A Fluid"

Transport Algorithm That Works", J. Comp. Phys. 11, 38, 1973.

47.

Zalesak, S.T., "Fully Multi-Dimensional Flux-Corrected Transport Algorithms for Fluids", J. Comp. Phys. 11, 38, 1973.

48.

Steger, J.L. and Warming, R.F., "Flux Vector Splitting.of the Inviscid
Gas Dynamic Equations with Application to Finite Difference Methods" J.,
Comp. Phys. 40, 263, 1981.

49.

Gresho, P.M. and Lee, R.L., "Don't Supress the Wiggles --They're Telling
You Something!", in Hughes, T.J.R. (ed.) Proc. of a Symp. on Finite
Element Meth. for Convection Dominated Flows, ASME Winter Annu. Meet.,
New York, 37, 1979.

50.

Heinrich, J.C.,
Huyakorn, P.S., Zienkiewicz, O.C. and Mitchell, A.R.,
"An 'Upwind 1 Finite Element Scheme for Two-Dimensional Convective ;
Transport Equation", Int. J. Num. Meth. Eng., 11, 131, 1977.

51.

Gottlieb, D. and Orszag, S.A., "Numerical Analysis of Spectral Methods",

NSF-CBMS Monograph, No. 26, Soc. Ind. and Appl. Math., Philadelphia, PA,
1977.
,

52.

Gottlieb, D., Hussaini, A.K. and Orszag, S.A., Paper Drafted for
Publication (private communication).

53.

Orszag, S.A., "Spectral Methods for Problems in Complex Geometries",

Numerical Methods for PDE's, Academic Press, 273, 1979.

54.

Hirt, C.W., Amsden, A.A. and Cook, J.L., "An Arbitrary

Lagrangian-Eulerian Computing Method for All Flow Speeds", J. Comp.
Phys. 14, 227, 1974.

55.

Thompson, J.F., Warsi, Z.U.A. and Mastin, C.W., "Boundary-Fitted

Coordinate Systems for Numerical Solution of Partial Differential
Equations- A Review", J. Comp. Phys. 47, 1, 1982.

-214-

56.

Dwyer, H.A., Kee, R.J. and Sanders, B.R., "An Adaptive Grid Method for
Problems in Fluid Mechanics and Heat Transfer", AIAA Computational Fluid
Dynamics Conference, Williamsburg, VA, 1979.

57.

Brackbill, J.V. and Saltzman, J.S., "Adaptive Zoning for Singular

Problems in Two Dimensions", J. Comp. Phys. 46, 342, 1982.

58.

Klopfer, G.H. and McRae, D.s., "The Non-Linear Modified Equation

Approach to Analyzing Finite Difference Schemes", AIAA Computational
Fluid Dynamics Conference, Palo Alto, CA, 1981.

59.

Chandra, R., "Conjugate Gradient Methods for Partial Differential

Equations", Ph.D. Thesis, Yale University, Dept. of Computer Science,
1978. .

60.

Concus, P., Golub, G.H. and O'Leary, D.P., "A Generalized Conjugate
Gradient Method for the Numerical Solution of Partial Differential
Equations", Stanford University, STAN-CS-76-5333, 1976.

61.

Settari, A. and Aziz, K., "A Generalization of the Additive Correction

Methods for the Iterative Solution of Matrix Equations", SIAM J.
Numerical Analysis, 10, 506, 1973.

62.

Hutchinson, B.R. and Raithby, G.D., "A Multigrid Method Based on the
Additive Correction Strategy", Numerical Heat Transfer, 9, 511, 1986.

63.

Brandt, A., "Multi-Level Adaptive Solutions to Boundary-Value Problems",

Math. Computation, 31, 333, 1977.

64.

Neely, G.M. and Claus, R.W., "Accelerated Convergence for Incompressible

Flow Calculations", NASA TM-86863, 1985.

65.

Caretto, L.S., Gosman, A.D., Patankar, S.V. and Spalding, D.B., "Two
Calculation Procedures for Steady Three-Dimensional Flows with
Recirculation", Proc. of Third Int. conf. on Numerical Methods in Fluid
Dynamics, 60, 1972.

66.

Srivastava, B.N., Werle, M.J. and Davis, R.T., "Viscous Shock Layer
Solutions for Hypersonic Sphere Cones", AIAA J. 16, 2, 136, 1978.

67.

Srivastava B.N., Werle, M.J. and Davis, R.T., "Numerical Solution of

Hypersonic Viscous Shock Layer Equations", AIAA J. 17, 1, 107, 1979.

68.

Keller, H.B. and Cebeci, T., "Accurate Numerical Methods for Boundary
Layer Flows- I. Two-Dimensional Laminar Flows-", in Holt, M. (ed) Proc of
2'nd. Inter. Conf. on Numerical Methods in Fluid Dynamics, New York, 1971.

69.

Leonard, B.P., "A Stable and Accurate Convective Modelling Procedure

Based on Quadratic Upstream Interpolation", Comp. Math. Appl. Mech.
Eng., 19, 59, 1979.

-215-

70.

Raw, M.J., Personal Communication, ASC, 1986.

71.

Wilkes, N.S. and Thompson, C.P., "An Evaluation of Higher Order Upwind
Differencing for Elliptic Flow Problems", Rept. CSS-137, AERG Harwell,
1983.

Coakley, T.J., "Numerical Method for Gas Dynamics Combining

Characteristic
and Conservation Concepts", AIAA Paper No. 81-1257,
14tn Fluid and Plasma Dynamics Conference, Palo Alto, CA, 1981,

72.

73.

Shyy, W., "Numerical Outflow Boundary Condition for Navier Stokes Flow
Calculations by a Line Iterative Method," AIAA J. 23, 12, 1847, 1985.

74.

Stubley, G.D., Raithby, G.D., Strong, A.B. and Woolner, A.K.,

"Simulation of Convection and Diffusion Processes by Standard Finite
Difference Schemes and by Influence Schemes", Comp. Methods Eng. Appl.
Mech., 35, 153, 1982.

75.

Schneider, G.E. and Raw, M.J., "A Skewed Positive Influence Coefficient
Upwinding Procedure for Control Volume Based Finite Element
Convection-Diffusion Computation", Numer. Heat Transfer, 8, 1, 1986.

76.

Collatz, L., The Numerical Treatment of Differential Equations, Springer

Verlag, Berlin, 1960.
-;

77.

Krause, E., Hirschel,

E.H. and Kordulla, W., "Fourth Order
'Mehrstellen 1 - Integration for Three-Dimensional Turbulent Boundary
Layers", Comp. Fluids 4, 77, 1976.

78.

Adams, Y., "A Hermitian Finite Difference Method for the Solution of
Parabolic Equations", Comp. Maths. Appls., 8, 393, 1975.

79.

Anderson, .0.A., Tannehill, J.C. and Pletcher, R.H., Computational Fluid

Mechanics and Heat Transfer, Series in Computational Methods in
Mechanics and Thermal Sciences, Hemisphere, 1984.

80.

Rubin, S.G. and Graves, R.A., "A Cubic Spline Approximation for Problems
in Fluid Mechanics", Tech. Report 74-T1, Old Dominion Univ., Norfolk,
VA, 1974.

81.

Swartz, B.K., "The Construction and Comparison of Finite Difference

Analogs of Some Finite Element Schemes," in Mathematical Aspects of
Finite Elements in Partial Differential Equations," (C.de Boor, ed.),
Academic Press, NY, 27,9, 1974.

82.

Ciment, M., Leventhal, S.H. and Weinberg, B.C., "The Operator Compact
Implicit Method for Parabolic Equations", J. Comp. Phys. 28, 125, 1978.

83.

Wong, H.H. and Raithby, G.D., "Improved Finite Difference Methods Based
on a Critical Evaluation of Approximation Errors", Numer. Heat Transfer,
2, 139, 1979.

-216-

84.

Lecointe, Y. and Piquet, J., "On the Use of Several Compact Methods for
the Study of Unsteady Incompressible Viscous Flow Round a Circular
Cylinder", Computers and Fluids, 12, 4, 225, 1980.

85.

Meijerink, J.A. and Van Der Vorst, H.A., "An Iterative Solution Method
for Linear Systems of which the Coefficient Matrix is a Symmetric
M-Matrix", Math, Comp. 1977.

86.

Dongarra J.J., Leaf, G.K. and Minkoff, M., "A Pre-conditioned Conjugate
Gradient Method for Solving a Class of Non-Symmetric Linear Systems",
ANL-81-71, Argonne National Laboratory,.Argonne, IL 1981.

87.

Hestenes, M. and Steifel, E., "Methods of Conjugate Gradients,for

Solving Linear Systems", J. Res. Nat. Bur. Standards 49, 409, 1952.

88.

Phillips, R.E. and Schmidt, F.W., "Multigrid Techniques for Numerical

Solution of the Diffusion Equation", Numer. Heat Transfer, 7, 251, 1984.

89.

Raithby, G.D. and Torrance, K.E., "Upstream-Weighted Differencing

Schemes and Their Application to Elliptic Problems Involving Fluid
Flow", Comp. Fluids, 8, 12, 191, 1.974.

90.

Smith, R.M. and Mutton, A.G., "The Numerical Treatment of Advection: A

Performance Comparison of Current Methods", Numer. Heat Transfer, 5,
439, 1982.

91.

Stubley, G.D., "A New Discrete Method for Convection Dominated Flows
Based on a Clear Understanding of Solution Errors", Ph.D. Thesis,
Unviersity of Waterloo, 1981.

92.

Burgraff, O.R., "Analytical and Numerical Studies of Structure of Steady

Separated Flow", J. Fluid Mech. 24, 2, 113, 1966.

93.

Vanka, S.P., Personal Communication, Argonne National Labs, 1986.

94.

Armaly, B.F. and Durst, F., "Reattachment Length and Recirculation

Regions Downstream of a Two-Dimensional Single Backward Facing Step",
Symp. on Momentum and Heat Transfer Processes in Recirculating Flows,
HTD, 13, ASME 101'st Winter Annual Meeting, Chicago, 1, 1980.

95.

Gh1a, U., Ghia, K.N. and Shin, C.T., "High-Re Solutions for
Incompressible Flow Using the Navier Stokes Equations and a Multi-Grid
Method", J. Comp. Phys. 48, 387, 1982.

96.

Khan, Z.A., McGuirk, J.J. and Whitelaw, J.H., "A Row of Jets in
Crossflow", Fluid Dynamics of Jets with Applications to V/STOL, AGARD
CP-308, Paper 14, 1982.

97.

Crabb, D., Durao, D.F.G. and Whitelaw, J.H., "A Round Jet Normal to a
Crossflow", J. Fluids Eng. 103, 1, 142, 1981.

-217-

APPENDIX A
IMPLEMENTATION OF MODIFICATIONS IN THE PRATT AND WHITNEY 3-D TEACH CODE TO
IMPROVE ACCURACY AND EFFICIENCY
INTRODUCTION

The following notes briefly outline the modifications of the Pratt and
Whitney 3-D TEACH code to incorporate various numerical techniques for
improving accuracy and computational efficiency. Relevant modifications for
the 2-D TEACH code follow a similar implementation structure and corresponding
details can be appreciated fully by a careful study of the material presented
here.
The outline presented includes descriptions of new routines and
modifications to the supplied TEACH code. These notes assume that the reader
is familiar with the supplied TEACH code and all relevant documentation,
as well as the appropriate discussion presented Section 8 regarding
formulation and extension of schemes.
MODIFICATIONS FOR IMPROVING EFFICIENCY

To improve the computational efficiency of TEACH, the following

numerical techniques were implemented:
i) Residual reduction convergence criterion for solution of the pressure
correction equation,
ii) Stone's Strongly Implicit Procedure (SIP) for the solution of the
pressure correction equation,
iii) Incomplete Choleski (1C) for the solution of the pressure correction
equation,
iv) Additive Correction Multigrid (ACM) to accelerate the convergence of SIP
and 1C. ACM was implemented for both 2x2x2 and 3x3x3 blocks,
v) Block Correction (BC) to accelerate the convergence of 1C with the
option of further accelerating convergence by ACM.
Originally, there was no intention to include BC in the modifications of
the 3-D TEACH code, Section (8.1), however, for some 3-D applications,
particularly involving geometries with high cell aspect ratios, the
convergence acceleration provided by ACM was limited. For these applications

-218-

BC provides considerably greater acceleration to convergence. To take

advantage of the features of both acceleration schemes, the two acceleration
schemes have been implemented so that they can be used together. As
demonstrated in Section (8.1), the combination of the two techniques results
in convergence acceleration which exceeds the acceleration provided by either
technique alone.
Residual Reduction Convergence Criterion
To implement the residual reduction convergence
criterion for .the
**
solution of the pressure correction equations, Section (4.3), modifications to
the subroutines CALCP and PISO were required. These changes include the
introduction of the following variables:
RES (I,J,K) = array of residuals for pressure correction equation
RESNOW =

variable containing the square root of the sum of the

squares of values of entries in RES of current p1
iteration.

RESMAX =

maximum value of resnow

RDCP

required reduction factor of residual, iteration

terminated when RSDNOW < RSDMAX*RDCP. Note: RDCP is an
input parameter passed into CALCP and PISO via the
labelled COMMON/PXLR8/.

The residual reduction convergence criterion is implemented in both pressure

correction stages of PISO (including the single correction stage of SIMPLE)
and is in effect only when SIP or 1C with or without acceleration is being
used.
Strongly Implicit Procedure
To implement Stone's SIP,
the following new subroutines were
provided:
i) ITRSOS with entry ITRSIS to set up the pointers into the real workspace
for subsequent calls to SIPO and SIPI. This data structure arrangement,
requiring the set up of pointers, was adopted to facilitate a modular
code structure.

-219-

ii) SIPO with entry SIPI to perform one SIP iteration. The SIPO portion of
the routine performs the approximate factorization of SIP required only
before the first SIP iteration and the SIPI portion performs the forward
and backward substitution required for each SIP iteration,
iii) RESIOL to determine residuals of the equations.
iv) NOTENF to generate an error message when there is not enough available
work space.
.. .
v) GUTTLE to generate a message indicating which routine is being used.
The argument list of these subroutines and entries include the following:
T =
V =
RES =
AP\
AE
AM
AN V
AS
AU
AD /
BP =
RLX =

dependent variable array

correction array to current values of T
residual array

coefficients of algebraic equations

source term of equation

relaxation = 1. + partial cancellation of Stone
1.0 < RLX < 2.0; for RLX set to unity, equivalent to 1C
(recommended use of 1C subroutines described below)
IOFLE = unit number for output
IB, IE, JB, JE, KB, KE = limits on index triplet (1,j,k)
ID, JO, KD = dimensions of arrays
WRK = work space containing NAVAIL elements
NAVAIL = number of available elements or words available in WRK
NREQ = number of elements WRK required by SIP.
Incomplete Choleski
The implementation of SIP with RLX=1 as described above is algebraically
(59)
equivalent to 1C.
However, the use of SIP routines for 1C results in an
unneccessary and considerable storage penalty. To overcome this, 1C was
implemented separately with the following new subroutines provided:

-220-

i) ITRSOI with entry ITRSII to set up the pointers into the real workspace
for subsequent calls to ICO and Id. Again, this data structure
arrangement requiring the set up of pointers was adopted to facilitate a
modular code structure.
ii) ICO with entry ICI to perform one 1C iteration. The ICO portion of the
routine performs the approximate factorization of 1C required only
before the first 1C iteration, and the ICI portion performs the forward
and backward substitution required for each 1C iteration. The argument
list of these subroutines and entries are identical to those for the
implementation of SIP.
Routines Providing No Acceleration of Base Solvers

Identifying SIP and 1C as the base solvers used to solve for pressure
corrections, routines designed to provide the desired convergence of the base
solvers are required. To provide no acceleration of the base solvers, the
following new subroutines are provided:
i) NOACO with entry NOACI to call the appropriate base solver routines for
approximate factorization and forward and backward substitution. The
base solver routines called depend on the value of RLX:
If RLX < 1.001, then 1C is used
IF RLX > 1.001, then SIP is used
The argument list of these subroutines and entries are identical to
those for the implementation of SIP.
Routines Providing BC Acceleration
To provide BC acceleration of the 1C base solver, the following new
subroutines are provided:
i) ITRBCO to set up pointers into workspace for subsequent call to BCD,
ii) ITRBCI to set up pointers into workspace for subsequent call to BCI,
iii) BCD to call COFBC to determine the coefficients of the block correction
equations for each of the three (E-W, N-S, and U-D) index directions and
to call ITRSOI to perform the approximate factorization for 1C,
iv) BCI to call SORBC.to determine source term of block correction equations
for each of the three index directions, to call FACTR1 and SWEEP1 to
solve the block correction fields, to call TMOOBC to apply the
appropriate block correction and to call ITRSII to perform the forward
and backward substitution of 1C,

-221-

v) COFBC to determine the coefficients of the block correction equations

for each of the three index directions,
vi) SORBC to determine the source terms of the block correction equations
for each of the three index directions,
vii) TMODBC to apply the block correction (note that only the block
correction that has the largest sum of the squares, of corrections is
applied),
viii) FACTR1 and entry SWEEP1 to perform Tri-Oiagonal Matrix Algorithm for
solving the three block correction fields.
In addition to the arguments described above, the argument lists of the
above routines include the following:
VEWB = E-W block correction vector
APEWB, AEB, AWB, BPEWB = coefficients of E-W block correction equations
VNSB = N-S block correction vector
APNSB, ANB, ASB, BPNSB = coefficients of N-S block correction equations
VUDB = U-D block correction vector
APUDB, AUB, AOB, BPUDB = coefficients of U-D block correction equations
Note: To ensure that block correction equations are not singular, one block
correction equation for each index direction is adjusted so that the
corresponding block correction is zero. Also, BC is currently implemented for
1C only. Implementation for SIP is readily accomplished by adding appropriate
calls to ITRSOS and IRSIS.
Routines Providing Additive Correction Multigrid Acceleration of Base Solvers
To provide ACM acceleration of the 1C and SIP base solvers as well as
IC-BC, the following new subroutines are provided:
i) XLR80M to call ACMPNT to set pointers into workspace, to call NCRMNT to
determine the number of fine grid control volumes in each coarse grid
block, to call COF6EN to determine coefficients of coarse grid additive
correction equations and to call ITRSOI, ITRSOS or LTRBCO to perform the
approximate factorization for all grids. The current implementation of
ACM required the introduction of a number of variables and vectors
including the following:
TC =
vector of pointers into real workspace pointing to first element
of coarse grid additive correction arrays
RESC= vector of pointers into real workspace pointing to first element
of coarse grid residual arrays

-222-

vc
we =

vector of pointers into real workspace pointing to first

element of coarse grid corrections to additive corrections,
required for current implementation of base solvers
vector of pointers into real workspace pointing to first
element of base solver workspaces required for each grid.

ARC
AEC
AWC
ANC
ASC
AUC

vector of pointers into real workspace pointing to first

elements of coefficients of additive correction equations for
each grid

AOC

BPC =

vector of pointers into real workspace pointing to first

element of coarse grid arrays of additive correction equation
source term
INC, JNC, KNC = vector of pointers into integer workspace pointing to
first elements of vectors of increments for each grid
IBC, IEC, JBC, JEC, KBC, KEC = vectors containing limits on index
triplet (i,j,k) for each grid.
IDC, JOC, KDC = vectors containing dimension of arrays for each grid
RESNOW = vector of square root of the sum of the squares of the current
residuals for each grid
RESPRV = vector of square root of the sum of the squares of the
previous iteration residuals for each grid
RESO = vector of square root of the sum of the squares of the
original residuals for each grid
IBLK, JBLK, KBLK = nominal size of blocks
RESRAT = required rate of convergence
RESRDC = required reduction of residuals
WRKMAX = maximum number of allowable total work units
NLVLS = number of coarse grids
LVL
= current grid level, finest grid is LVL = 0, coarsest grid is
LVL = NLVLS

11)

Entry XLR8IM to call SORGEN to determine source terms of coarse grid

additive correction equations, to call ITRSII, ITRSIS, or ITRBCI and
TCCOR to perform appropriate forward and backward substitution and

-223-

obtain required update of solution for the current grid, to call

RESID1, RESCAL and LEVCON to evaluate residuals on current grid, to
evaluate sum of squares of residuals and decide whether to remain on
current grid or to" move to finer or coarser and to call TMODAC to apply
additive correction to finer grid solution and adjust finer grid
residuals,
iii) ACMPNT to set up pointers into workspace for coefficients of all
coarser grid additive correction equations. The number of coarse grids
is based on logic used to determine increments, see (iv).
iv) NCRMNT to determine increment of fine grid control volumes which are
used to make up a coarse grid block in each of three index directions.
Increments in each direction are nominally set to value of IJKBLK,
except at boundaries where the number of fine grid control volumes in
an index direction is not evenly divisible by IJKBLK. In this case the
remaining increment is set to the integer remainder of the division.
The logic of NCRMNT is set up so that the number of blocks in any index
direction is not less than 2. The number of coarse grid levels assumed
in ACMPNT is such that the coarsest grid is 2x2x2.
v) COFGEN to determine coefficients of coarser grid additive correction
equations. The coarse grid AP coefficient is determined by first
summing fine grid AP values over coarse grid blocks. Then coarse grid
AP values are modified by subtracting fine grid AE, AW, AN, AS, AU and
AD coefficients where these coefficients are not coincident with coarse
grid control volume faces. The coarse grid coefficients AE, AW, AN, AD
are determined by summing the fine grid coefficients, where fine grid
and coarse grid control volume faces are coincident.
vi) SORGEN to determine the source term of coarse grid additive correction
equations. These source terms are simply determined from the sum of
fine grid residuals over coarse grid blocks.
vii) RESCAL to calculate the square root of the sum of the squares of the
residuals.
viii) LEVCON to set a flag controlling movement of ACM from one grid to
another. Note that the current implementation of ACM adopts the
flexible cycle multigrid approach. ACM moves to a finer grid if
residuals are 20 percent of original value and moves to a coarser grid

-224-

if residuals have not been reduced to 10 percent of the previous value.

ACM also moves to a finer grid if the number of total work units, summed
over all grids but the LVL = 0 grid, exceeds 1.0.
ix) TCCOR to apply corrections from base solver to the current estimate of
solution.
x) TMOOAC to apply additive correction to finer grid solution and adjust
finer grid residuals. Fine grid residual adjustments are made in
several steps: the first step is to adjust all fine grid residuals
assuming that the additive corrections to neighbouring fine grid control
volumes are the same; the second step then adjusts all fine grid
residuals where fine grid faces and coarse grid faces are coincident
(where neighbouring additive corrections are not necessarily the same).
In addition to the arguments already described, the argument lists of
the ACM routines also include the following:
dependent variable array for coarse grid
T8
residual array for coarse grid
RESB
\
APB
AEB
AWB
ANB =
ASB
AUB
ADB
BB / =

coefficients of coarse grid additive correction equations

source term of coarse grid additive correction equation

LB.LE.MB.ME.NB.NE = limits on index triplet (l.m.n) for coarse grid
LD, MD, ND = dimensions of arrays for coarse grid.
Modifications to Routines to Calculate Pressure Corrections
To invoke the various improved techniques for solving the equation for
pressure correction, a number of calls were added to the CALCP and PISO
routines. These calls can be broken into two types; calls to routines with
names beginning with XLR80 which perform to necesssary initializations and
approximate factorization and calls to routines with names beginning with
XLR8I to perform one base solver iteration with acceleration. In general,
calls to XLR80 routines are required only when the A coefficients of the p1
equation change. Therefore a call to XLR80 1s not required prior to solving

-225-

the second set of p1 of PISO. The call to the appropriate XLR80-XLR8I-pair is

:
controlled by the value of IPXLR8:
.
IPXLR8 = 0 use original Alternating Line Gauss-Seidel
= 1 no acceleration of SIP or 1C
= 2 ACM acceleration of SIP or 1C (2x2x2 blocks)
= 3 ACM acceleration of SIP or 1C (3x3x3 blocks)
= 4 BC acceleration of 1C
= 5 ACM acceleration of IC-BC (2x2x2 blocks)
. '
= 6 ACM acceleration of IC-BC (3x3x3 blocks)
For IPXLR8 = 1,2 or 3 the choice of SIP or 1C is dependent on the value of
RLXP, the parameter controlling the partial cancellation of SIP
RLXP < 0.001 : 1C is used
> 0.001 : SIP is used
The variables IPXLR8, RLXP and ROCP are passed into CALCP and PISO via the
labelled COMMON/PXLR8/. Recommended values for these variables are:
IPXLR8 = 5 (IC-BC-ACM 2x2x2 blocks)
RLXP = 0.0 (1C, by default if IPXLR8 = 5)
ROCP = 0.1 (terminate p1 solution if p1 residuals are reduced to within
10 percent of their original value)
Although these values are not necessarily optimal, they should be
appropriate for a wide range of problems.
.;
For the solution of the pressure correction equations required in CALCP
and PISO, the coefficients and source terms of pressure correction equations
must have effects of all boundaries eliminated. This is the default for the
TEACH code. As a result no additional modifications are required;
Finally, the modified PISO routine contains changes required to correct
the block-off logic when applying pressure corrections to the velocities.
MODIFICATIONS FOR IMPROVING ACCURACY

To improve the accuracy of the TEACH code the Linear Profile (LP) and
Mass Weighted (MW) Skewed Upstream Differencing Schemes (SUDS) were
implemented. Details of the implementation for these techniques can be found
in Section (8.2). To implement MW-SUDS and LP-SUDS, new subroutines were
required to determine integration point expressions, to evaluate positive
octant mass flows and to assemble integration point influences. Modifications
to existing routines to calculate exchange coefficients for various variables
are also required.

-226-

Integration Point Expression for MW-SUDS and LP-SUDS

To determine the Integration point expression, subroutine IPEQ is
provided. The routine is written using the notation outlined in Section (8.2)
and determines the integration point expression for only one Integration
point. Oil. IPEQ is broken into four sections: i) preliminaries, ii) LP-SUDS,
iii), MW-SUDS, and iv) Physical Advection Correction (PAC).
In the preliminary section of IPEQ, the exponential weighting factor and
the value of pV/L are determined. To determine the value of pV/L the first
step is to calculate the six different values of pV/L corresponding to the
cases where the streamline intersects each of the six sides of the flux
element. The value of pV/L chosen is that which corresponds to the smallest
positive value of L (i.e., the side first intersected by following the
streamline upstream) or the largest value of pV/L.
The LP-SUDS section of IPEQ determines the coefficients of the
integration point equation from a bilinear interpolation of the nodes on the
flux element side intersected by the streamline. For low grid Peclet numbers
LP-SUDS is blended with a trilinear interpolation of the eight flux element
nodes. The blending is determined from the exponential weighting factor, B.
The MW-SUDS section applies the positive mass weighted algorithm, to
determine coefficients of the integration point equation. The mass weighted
algorithm is applied to the 000, 010, 001 and Oil set of nodes if the
mass flow at the X011 integration point is positive and to the 100, 110, 101,
and 111 set of nodes if the mass flow is negative. Note that the notation
used for the positive mass flows of octant faces is slightly different in IPEQ
than that used in Section (8.2). In the IPEQ routine mass flows through
integration point faces and faces adjacent to nodes are required on an octant
basis. See description of OCTMAS below. For low grid Peclet numbers MW-SUDS
is blended with a linear interpolation of the flux element nodes Oil and 111.
The physical advection correction is determined from a trilinear interpolation of nodal estimates of the correction.
The argument list of IPEQ includes the following:
AIP
BIP
ANPOOO
ANP100 %. = coefficients of integration point equations

ANP111

-227-

0X0,DX1

DYO.DY1

DZO, DZ1 =
RVXOOO, RVX010,..
RVYOOO, RVY100,..

distances in X direction from integration

flux element sides
distances in Y direction from integration
flux element sides
distances in Z direction from integration
flux element sides
= mass flux at integration point on faces
coincident with X plane
,
= mass flux at integration point on faces

point to
point; to_
point to

coincident with Y plane

= mass flux at integration point on faces

coincident with Z plane
FNPOOO, FNP100... = positive (into octant) mass flows
through faces adjacent to flux element node
FXIOOO, FXI100...
= positive mass flows through octant
integration point faces coincident with X plane
FYIOOO, FYI100,.., = positive mass flows through octantintegration point faces coincident with Y plane
FZIOOO, FZI100,.., = positive mass flows through octant
integration point faces coincident with Z plane
QNPOOO, QNP100,... = estimates of PAC at flux element nodes
RVZOOO, RVZ100...

DIFIP
ACC

= diffusion coefficient at X011 integration point

= parameter controlling formulation of integration

point equation
= 0 then MW-SUOS without PAC
= r then MW-SUDS with PAC
=2 then LP-SUOS with PAC

Positive Octant Mass Flow

MW-SUOS for three-dimensional applications requires the evaluation of

positive mass flows through octant faces. The routine OCTMAS performs this
function. Mass flows are defined to be positive, if they are into the
octant. For flows leaving the octant, the value of the positive mass flow
values are required for four sets of faces: 1) flows through the X integration
point face of the octant, ii) flow through the Y integration point face of the
octant, iii) flow through the Z integration point face of the octant, and iv)
net flow through octant faces intersecting at octant node.

-228-

In addition the the variables described above,

OCTMAS includes the following:
AXOOO, AX010,... = area of X integration point
AYOOO, AY100,... = area of Y integration point
AZOOO, AZ100,... = area of Z integration point

the argument list of

faces of octant
faces of octant
faces of octant

Flux Element Assembly

To assemble the influences of integration point equations of a flux
element into the appropriate control volume flux balance equation, the routine
FEASM is provided. There are four major sections to FEASM: i) data statements
initializing maps, ii) determination of positive mass flows, 1ii)
determination of integration point equation coefficients for each integration
point of the flux element, and iv) assembly of influences.
As described above, the routine IPEQ, provided to determine the
coefficients of the integration point equations, is written for only the X011
integration point. To make use of one routine for all twelve integration
points of a flux element it is necessary to use transformations.

In FEASM

these transformations are made through the use of maps. To use these maps it
is first necessary to introduce a change in notation.

Instead of the indexed

triplet notation described in Section (8.2) and used in IPEQ and OCTMAS, node
points and integration points are assigned the integer numbers listed below:
Index Triplet
Numbered
Node Notation
000

100

010

110

001

101

011

111

-229-

Integration Point
Notation

XOOO
X010
X001
X011

1
2
3
4

YOOO

YIOO

Y001
Y101
ZOOO
Z100
Z010
Z110

7
8
9
10
11
12

. ,

* :

'

. , .,

Using the numbered notation, transformations are readily created and

implemented. For instance, if the coefficients of the integration point for
integration point 1 (or integration point X001) are required, then IPEQ can be
called with integration point 1 arguments used in place of integration point 4
arguments, integration point 2 arguments in place of integration point 3
arguments, etc. The transformation of integration points is given in FEASM in
the map IM. Note that a map is required for each integration point with 12
entries in each map, hence IM is dimensioned 12x12. In addition to the
integration point maps, node point maps are required. For each of the twelve
integration points,
the eight entries required to map nodes are given by NM.
>
Maps are also provided for three dimensions of the flux element and the sense
of the coordinates under each of the twelve transformations. The map for the
flux element dimensions assumes that dimensions have been assigned numbers in
the following order:
Index Notation
Numbered Notation
DXO
DX1

1
2

DYO
DY1
DZO
OZ1

3
4
5
6

-230-

The assignment of indexed notation variables to numbered notation

variables follows the call to OCTMAS. This is followed by twelve calls to
IPEQ. After normalizing the coefficients of the integration point equations
and multiplying by the appropriate mass flows, the flux element influences are
assembled. This assembly is done on a flux element octant basis. Each
control volume octant has three flux element integration points with each
integration point expressed in terms of the eight flux element nodes.
In addition to the variables described previously, the argument list of
FEASM includes the following:
ADSW
ADS
AOSE

exchange coefficient arrays

AP

AUNE
BT = source term coefficient
I,J,K = integer index triplet of 000 flux element mode
ID.JD.KD = dimensions of all exchange coefficient arrays
Modifications to Routines Calculating Exchange Coefficients
MW-SUOS and LP-SUDS discretizations are implemented for the calculation
of the coefficients of equations for u (CALCU), v (CALCV), w (CALCW),
turbulent kinetic energy (CALCTE), and dissipation (CALCEO). In general, the
type of discretization used is controlled by the value of IFESKW passed into
the routines through the labelled COMMON/FESKW/:
IFESKW = 0 default to baseline TEACH
= 1 MW-SUOS
= 2 LP-SUDS
As outlined above, the flux element assembly of coefficients using SUDS,
invoked by calling FEASM, requires information regarding geometry, mass fluxes
and estimates of diffusion and source term influences. Geometric information
is readily available in TEACH. The mass flux terms are required on an octant
basis and re-derived from the control volume mass fluxes which are stored in
RVX, RVY and RVZ and passed as arrays through /FESKW/. To obtain diffusion
and source term influences when using SUDS, the TEACH calculations of diffusion

-231-

calculations of diffusion coefficients and source terms are retained,and

assembled in the A and SU coefficients prior to the call to FEASM. Note
however, that no convective influences are included until after the call to
FEASM. The results is that, before the call to FEASM, the A coefficients
contain all the sufficient diffusion information and SU coefficients contain
source term information. The estimate of combined diffusion and source term
influences, required to estimate the physical advection correction term, is
obtained from the residual of the conservation equation (without convective
influences) using the best available estimates of the dependent variable.
Normalized by the volume of the control volume, these influences are stored in
the array QU passed into the subroutine through /FESKW/.
In the vicinity of boundaries, some modification of the procedure
outlined above is required. First, in TEACH, fictitious modes are located
coincident with the boundary. The implementation of flux element SUDS retains
this by appropriately adjusting distances between integration points and flux
element boundaries. Second, in the supplied TEACH, any influence.of any
boundary node which is not E,W,N,S,B or F must be zero. In the flux element
implementation of SUDS this assumption remains valid if, in flux elements
adjacent to boundaries, only MW-SUDS is used. This assumes that, at specified
openings, the components of velocity tangent to the boundary are zero.
Finally, at boundary nodes the value of QU cannot be defined in the manner
given above. Instead, boundary QU values are set equal to adjacent interior
values.
With the necessary geometric, mass and diffusive fluxes and source term
information available, coefficients are assembled on a flux element basis by
calling FEASM. The evaluation of mass fluxes and source term influences,
required by FEASM, are determined in a manner which is appropriate for either
Cartesian or cylindrical coordinates.
Finally, for kinetic energy and dissipation equations the value of QU
are everywhere set to zero and, if IFESKW > 0, then only MW-SUDS, ACC = 0,
is used. This is done to ensure the required boundedness of the kinetic
energy and dissipation solutions.
MODIFICATIONS TO INPUT AND OUTPUT ROUTINES

To implement the modifications in TEACH for improving accuracy and

efficiency, it was necessary to add several new variables which must be
supplied by the user from an input f.ile. This required the modification of
the INPUT routine. These changes are summarized below:

-232-

i)
ii)
111)
1v)

v)
v1)

the labelled COMMONS /PXLR8/ and /FESKW/ are included

following card type 9, a statement to read FESKW (format
statement 5491, 15) was Introduced
following card type 11, a statement to read IPXLR8, RLXP, ROCP
(format statement 5471, 15, 2f5.3) was Introduced
the introduction of a statement to overlde the use of MW-SUDS or
LP-SUDS if BSUOS1 or BSUDS2 were specified, ie., MW-SUDS or
LP-SUOS are used only if ISKEW > 0 and FESKEW > 0
the introduction of a statement to ensure that RLXP = 0 for
IPXLR8 > 3. This ensures that only 1C is used with BC.
write and format statements to echo Input of changes 2 and 3
were introduced.

The routine XINPUT was also modified so that explanations of the

additional options could be provided. Note, all additional input is read from
unit 5 and additional output is directed to unit 6.
In addition to the changes listed above, the following modifications to
the BLOCK DATA and RESTRT as well as the MAIN were required:
1)
in the interest of saving compute resources (disk space and
memory) the inclusion of SKW3.INCL and Initialization of the
variable SMALL were removed from BLOCK DATA. The variable SMALL
is initialized in MAIN.
ii)
so that the restart feature of TEACH could be used, the variable
KMBSTR was removed from the second record read from unit 1 in
RESTRT.

-233-

APPENDIX B
DERIVATION OF COEFFICIENTS FOR R AND Q MATRICES IN THE COCI SCHEMB

It is shown in Appendix C that the truncation error defined by:

= [R]

- [Q]

(B.I)

can be expressed as, using Taylor Series Expansion:

= T
T *. + TT *.

2
3 3
44
T *, + TT *. +T T *. +
T2

where * = d */dx and

T = r[r6 + rP+ rW,]

zp
T

T = h [r- r 2zP

2F

Sr

(B.2)

6F
3zf
U3
- 3!
^T [re- rw - - (qe- qw) - -
h
12
4zP.

T-*T
V.

(-D"rw-t

(-1)'

for v = 3, 4, 5 and z =

-234-

M-H

(-D1

COCI 1s said to be formally fourth order if r in Equation (B.I) =

0(h ) implying that Tw = 0(h4) in equations (B.2). Since T5 = 0(h5),
then to obtain a fourth order scheme consideration need only be given to Tw
0 1 2
for v = 0,1,2,3 and 4. Rearrangement ofT =T =T =0 defines tridiagonal matrix operator R in terms of Q in equation (6.65), as:
4

2rw = -f

[q w (2-3zu i _ 1 ) + q p (2-zu.) + q e (2+zu ui )]

2re = -|

[q^Z-zu^^ + q p (2+zu i ) - q e (2+3zu ui )]

e
w
= - (r + r )

(B.3a)

(B.3b)

(B.3c)

COCI is uniquely defined by further insisting that T3=T 4=0. The resulting
3
4
expressions for T and T become:
p
u
<.i P u
dp
w
re - rw - -f- (q e -q w ) - f- (q u. .,+qV ,) = 0
1 +l
1H
^
^

12F
re - rw - f

4zF
(q e -q W )- f

(qeuui - q^.-,) = 0

(B.4a)

(B.4b)

Normalizing Equations ( B . 3 ) and ( B . 4 ) by qP and rearranging,

^e _ AW

r
=

_| . t jw ( z u
h

, +

zu

+q e (zu i + 1 )]

2P
r e + A r w = -f- [q^l-zu^^ * 1 + q e ( U z u ) J

A6 _ AW

-235-

(B.5a)

(B.5b)

where
/si
. p ,x\1
. p .i = e, p, w
q = qi AT
r = ri AT,

Equations (B.5) can now be solved f o r r 6 , ,

and

to yield

= [6 + ( 2 u - 5 u ) z - u . u ] / [60
q = [6 + (Su.. - 2u^_1)z - u..u.j_.|] /
Based on the common denominators of q and'q' , a logical selection for
q is:
qp = 60 + I6(u1+l - u.^Jz - Au.^u^z2

(B.7a)

Hence, the remaining coefficients are:

q6 = 6 + (5ui - 2u1_l)z

(B.7b)

qW = 6 - (5u. - 2 u^-,)2

r =

"&
2lt2

(B.7c)

1-1

p
e
rW = ^,
[qw(2-3zu.
1-1.)+ q (2-zu.)
i + q (2 + zu,^)]
1+1
2hz

rp = - (re + rw)

''.

The above relationships provide the coefficients in the postulated COCI

expression, i.e.,

w ox

or [ R ] 0 = [0]

-236-

px

e px

(B'7e)
(B.7f)

APPENDIX C
DETAILS OF TRUNCATION ERROR SERIES FOR COCI DISCRETIZATION

The COCI scheme postulates that a tr1-d1agonal relation exists between

0.I and the spatial operator . 1 = 0r a.
+1 u.1 0. (assuming r 0 = constant
1
w

"

and u varies in space) of the form:

w
p_
e_
r 0_

+ r"

'

w/>x

+ r 0.

px

= q _

+ q^JC

. ,.

epx

q.JL

(C.I)

define as:
If the truncation error is noww defined

'w

where *^ and L? are the exact solution to equation (C.I). By expanding each neighbouring value of 0 andx in equation (C.I) about x.t the
following relationship results:
w
W

'

'

2
h

"

">
h

"

"\
h

<

4
h

'''

iv

+h

' ' '

"

iw

~ "

..... ] * r
n

(C.2)

.4

Collecting terms common to each derivative of 0, T is expressed as:

0

1
j

'
^

?
i

-237-

"
^

II I

T^

.^

^ I*

(C.3)

where it follows directly from equation (C.2) that

T0

, re

t rw t rp

T1 = h[re-

U2

T -

rw2F
e

r[r + r - -

tr M-,,V

(C.4)

3zr

er

(,'-,")-

- ^Ijjf <,*<-!>">-^

for v = 3, 4, 5 and z = h/F

It is important to note that the above deviation is equally valid,

ing for
allowing
for spatial variation of r0 and h, which was not accounted for here
only for simplicity.

-238-

NASA

Report Documentation Page

National -Aeronautics and

Space Administration

3. Recipient's Catalog No.

2. Government Accession No.

1. Report No.

NASA CR-180852
5. Report Date

4. Title and Subtitle

July 1988

Improved Numerical Methods for Turbulent Viscous

Redrculatlng Flows

6. Performing Organization Code

8. Performing Organization Report No.

7. Author(s)

None

A. Turan and J.P. VanDoormaal

10. Work Unit No.

533-04-11

9. Performing Organization Name and Address

11. Contract or Grant No.

Avco Research Laboratory, Inc.

2385 Revere Beach Parkway
Everett, Massachusetts 02149

NAS3-24351
13. Type of Report and Period Covered

Contractor Report
Final

12. Sponsoring Agency Name and Address

National Aeronautics and Space Administration

Lewis Research Center
Cleveland, Ohio 44135-3191

14. Sponsoring Agency Code

15. Supplementary Notes

Project Manager, David A. Jacqmln, Internal Fluid Mechanics Division, NASA Lewis
Research Center.
16. Abstract

The performance of discrete methods for the prediction of fluid flows can be
enhanced by Improving the convergence rate of solvers and by Increasing the accuracy of the discrete representation of the equations of motion. This report
evaluates the gains 1n solver performance that are available when various acceleration methods are applied. Various discretizations are also examined and two
are recommended because of their accuracy and robustness. Insertion of the
Improved discretization and solver accelerator Into a TEACH code, that has been
widely applied to combustor flows, Illustrates the substantial gains that can be
achieved.

17. Key Words (Suggested by Author(s))

18. Distribution Statement

Numerical method; Redrculatlng flows;

High order upwind; Differencing schemes;
Mult1gr1d algorithms
19. Security Classif. (of this report)

Unclassified

NASA FORM 1626 OCT 86

20. Security Classif. (of this page)

Unclassified - Unlimited
Subject Category 02

Unclassified

21. No of

"For sale by the National Technical Information Service, Springfield, Virginia 22161

22. Price*

National Aeronautics and

Space Administration

FOURTH CLASS MAIL

Lewis Research Center

Cleveland, Ohio 44135

ADDRESS CORRECTION REQUESTED

Official Business
Penalty (or Private Use S300

IM/NSA

Postage and Fees Paid

National Aeronautics and
Space Administration
NASA-451